Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 15, Article number: 18828 (2025) Cite this article
1162 Accesses
2 Altmetric
Metrics details
The emergence of digital technology has led to a significant increase in the importance of educational credential storage, exchange, and verification for organisations, enterprises, and universities. Academic record forgery, record misuse, credential data tampering, time-consuming verification procedures, ownership and control difficulties, and other problems plague the education sector. Machine learning (ML) and blockchain, two of the most disruptive methods, have replaced traditional techniques in the education sector with highly technological and efficient ways. Our study aims to propose a novel electronic educational document management technique using a blockchain-based fuzzy feed-forward convolutional temporal neural network that detects malicious users. Here, the training is carried out based on NLP analysis in document word weight indexing. This document management access control is based on role-based access with simulated remora swarm optimisation. In order to identify malicious users, this suggested system logs access requests on the blockchain and authenticated users. The findings demonstrate that this suggested architecture performs as intended in every case. The experimental analysis is based on a malicious user detection dataset regarding Prediction accuracy, Mean average precision, F-measure, Latency, QoS, Contract execution time, and Throughput. Based on dataset feature analysis, the proposed B-FCTNN_SRSO achieved a prediction accuracy of 98%, a mean average precision (MAP) of 95%, and an F1 score of 97%, with a latency of 96%. Additionally, based on blockchain security analysis, the B-FCTNN_SRSO attained a QoS of 97%, a precision of 94%, and a throughput of 96%.
Systems for managing electronic documents are employed in many different sectors. Specifically, electronic document management systems are essential for streamlining government agency paperwork procedures and civil servants’ work by enabling easy, efficient access to documents. These systems also automate repetitive tasks like tracking down relevant information, searching for it, and creating reports on the flow of documents. Nonetheless, public authorities process a vast number of documents annually within set processing times, and the efficacy and efficiency of public authorities are primarily dependent on the calibre and productivity of document exchanges¹. The daily requests may increase to several thousand as e-government advances. Intelligent algorithms will be more successful in government structures because documents and processes are stereotyped instead of systems with intricate and distinct organisational frameworks. In addition to preventing human error, machine learning may expedite processing of documents and prepare all data required for human decision-making. The filing cabinet’s development at the close of the nineteenth century marked the beginning of document management history. Edwin Granville Seibels created a vertical filing method in 1898 that arranges paper documents into boxes kept inside folded cabinets. For most of the 20 th century, these cabinets would remain dominant means of document storage in corporate sector². A collection of completed documents that users have downloaded, such as books, dissertations, conference papers, newspapers, full-text journals, and other database contents, is known as educational resource data. Many materials are gathered in commercial databases with permission to use for data expression and intellectual property protection in educational resources; resources from libraries are also searched, downloaded, integrated into specific databases³. These resources can also be downloaded, copied, distributed, and have other features that have led to problems with intellectual property. ML educational resource data may be categorised into four categories based on various types of educational resources: corporate, personal, government, and other public institutions. Due to data transfer barriers, educational resources may lose control of educational secret data in an ML environment, even though educational secret data is protected by a data backup system⁴. Artificial intelligence (AI), encompassing ML and DL, is widely seen as a game-changer in numerous industries and sectors, including manufacturing, advertising, healthcare, telecommunication^5,38, construction, and transportation. Since AI enables students to approach learning obstacles in a way that is customised to their unique experiences and interests, it will play a bigger role in higher education. AI-based digital learning techniques may adapt to each student’s knowledge level, preferred learning style, and learning objectives to help them get the most out of their education. In order to identify students’ areas of weakness and suggest courses that will improve their customised learning experience, it can also look at their prior academic records. In addition, teachers in higher education can devote more of their time to teaching and research by using AI to cut down on the time required for regular administrative activities⁶.
To introduce a novel blockchain-based fuzzy feed-forward convolutional temporal neural network (B-FCTNN_SRSO) for secure educational document management and malicious user detection.
To implement role-based access control using simulated remora swarm optimization (SRSO) for efficient document access management.
To use natural language processing (NLP) for document word weight indexing to improve document management efficiency.
To evaluate the proposed system using a malicious user detection dataset based on various performance metrics such as: Prediction accuracy: 98%, Mean Average Precision (MAP): 95%, F1-Score: 97%, Latency: 96%, Quality of Service (QoS): 97%, Precision: 94%, Throughput: 96%.
While some studies have attempted to predict student achievement, others have also classified educational data. Unal et al.⁷, focused on two sides of undergraduate students’ performance utilising DM approaches. The first step is to predict pupils’ academic performance at the conclusion of a four-year study programme. The second involves looking at how kids are developing and combining that with the results of predictions. He split the pupils into groups based on their levels of achievement: low achievement and high achievement. According to his findings, teachers must concentrate on a select group of courses that show especially strong or weak performance to provide timely warnings, assist underachievers, and provide guidance and opportunities for high achievers. Zhou & Huang⁸ used sixteen demographic variables, including age, gender, number of courses taken, internet connection, computer ownership, and attendance in class, to predict students’ academic success⁹. Among the ML methods, random forest, logistic regression, k-nearest neighbours, and SVM were able to predict students’ performance with prediction accuracy ranging from 50 to 81%. Heidari et al.¹⁰, created a model based on the students’ demographics and the grades they received for their in-term activities. In that study, classification methods based on Gradient Boosting Machine (GBM) were utilised to predict students’ academic progress. Findings indicated that nonattendance and achievement scores from the prior year were the best factors to use when estimating achievement scores. The authors discovered that demographic details like age, school, and neighbourhood may also be used to predict success or failure. The author presents a preliminary study about the creation, application, and delivery of LMS¹¹. An overview of learning analytics is given in the paper to help combine learning with data. According to study’s findings, learning analytical methods are most prominent methods in the literature. The four processes involved in creating such models are gathering relevant data, reporting, forecasting, acting, and fine-tuning the learning environment in response to the data. This study does not cover specific machine-learning algorithms that perform well with the model. Likewise, Dewangan & Chandrakar¹² summarises educational data mining by reviewing this area’s main ideas. Both studies summarised and explained the existing learning analytics and the subject of educational data mining and its methods, deviating from the systematic literature review requirements.
In addition, Wu¹³ offered an overview of educational data mining in another thoughtful literature review study. Rajendran et al.¹⁴ used input data like gender, wealth, board marks, and attendance to forecast students’ performances utilising ML algorithms like C4.5, sequential minimum optimisation (SMO), Naïve Bayes, 1-NN (1-Nearest Neighbourhood), MLP (multi-layer perceptron). After implementing correlation-based feature selection (CBFS) strategies to enhance method performance, they discovered that SMO outperforms other approaches in terms of effective average testing prediction accuracy, coming in at 66%. The author used artificial neural networks (ANNs)^15,55 to forecast student performance. When these methods were used with input characteristics, including grades, study periods, and test scores, they attained a high prediction accuracy of 85%. Razak et al., identified at-risk students before the next course, the author employed logistic regression, SVMs, decision trees (DTs), artificial neural networks (ANNs), NB classifier (NBC)¹⁶. Input components from an offline course were employed in this study, including grades, attendance, quizzes, weekly homework, team participation, project milestones, mathematical modelling activities, and exams. According to an analysis of data, NBC algorithm produced predictions with acceptable accuracy (85%). A study by Zhang¹⁷ employed ML methods to forecast students’ academic achievement in engineering courses. Exam results were the study’s output variable, and course grades from every semester were among the input features. The researchers found that while multilinear regression is useful for predicting success of every student in a course, support vector machines (SVMs) are better suited for predicting a single student’s performance. Rajendran et al.¹⁸, researched the most effective classifier to use social and personal input variables to predict students’ success in higher education. Through analysis of logs generated while students were using computers, certain probabilistic models—such as Bayesian knowledge tracing—have been utilised to forecast students’ performance. These models, however, are unable to forecast pupils’ latent tendencies. In a comparable setting, Navimipour et al.¹⁹, developed a hybrid adaption system that groups students according to commonalities and suggests the best learning materials for each group. To construct learner profiles, this system considers the users’ past activities, learning preferences, and knowledge levels. Next, the Nearest Neighbour algorithm (KNN) is used to group learners. As a result, it offers adjustments based on the characteristics of the acquired learner group rather than on an individual basis. Fahd et al.²⁰, used reinforcement learning-based adaption approach. All that this system needs to adjust and recommend is a learning path to meet the demands of the learners, which is their learning style. Similarly, Shi et al.²¹, suggested an adaptive e-learning method architecture based on reinforcement learning and a multi-agent system technique to suggest an adaptive learning path for a student who fits the following profile: verbal learning style, hearing impairment, and intermediate knowledge level.
In addition to structured data, there is a significant amount of unstructured and semi-structured data related to educational resources. Structured data is defined as having a set format and being of a specific length. Data without a set format and variable length are called unstructured data. Early on, copyright was typically used to protect data expression and products of machine learning-based educational resources. However, copyright only covers expressing ideas and data, not the data itself. It also covered the work’s selection, arrangement, system, and structure. Additionally, the data are made available for public dissemination. In that case, a more significant number of persons will unavoidably come into contact with them, and the service provider cannot ensure that the data’s intended use will adhere to legal requirements, hence raising the possibility of copyright infringement. In this scenario, the user and data operator enter into a legally binding agreement through a licence agreement for data work about purchasing a copyright licence for data work. The proposed blockchain-based machine learning model in data analysis is shown in Fig. 1.
Proposed blockchain based machine learning model.
This data set undergoes data cleaning to minimise noise and missing values. After removing the missing records, the data set is reduced to 133. Data set originally contained 17 missing values in various aspects from 150 records. There are 48 females and 85 males in the data set. The stage ID consists of 22 high level, 47 middle level, and 64 lower level. In addition, students are divided into three sections: section A has 69 students, section B has 49 kids, and section C has 15 students. One hundred eleven students have their father as their contact person, and 22 students have their mother as their contact person.
For word embedding models, we use the word2vec framework for training. The framework implements two distinct models and training methodologies. The first method, the Continuous Bag-Of-Words (CBOW) technique, attempts to anticipate a word by using its context—the surrounding words—as input. The other approach, the skip-gram method, guesses a word’s context using the word itself as input. Figure 2 shows a graphical representation of both approaches, with t representing the current word’s location and k representing the context window’s size. While the skip-gram approach works better for infrequent words, the CBOW model is faster overall. CBOW is less accurate in predicting unusual words because it averages the context word vectors to forecast the centre word. We have two sentences: “The food was devine” and “The food was delicious.” CBOW predicts words of interest based on context. Now that we want to predict the final word in the context [the food, was], the model is far more likely to suggest “delicious” because CBOW predicts the most likely word.
Continuous Bag-of-Words and skip-gram methods.
The word embedding algorithm is first algorithm. It is preferable to use this approach for rare words. The negative sampling algorithm is second training algorithm. Framework contains a large number of parameters. Consider simply the most important parameters. First, the dimensionality parameter finds number of dimensions of word vectors; generally speaking, a greater dimensionality is preferable. Nevertheless, the computing time increases with the number of dimensions. Next, depending on how close a word is to another word, the word2vec context window size parameter calculates the number of words that make up the word’s context. Moreover, minimum word frequency parameter determines how frequently a word appears in corpus to be considered. CBOW architecture trains method, centre word vector is found by taking the mean of the context vectors. Lastly, the noise word parameter is set to three and negative sampling is employed.
Algorithm 1Domain Words Extraction
search surface g (x, y, z) and template f (x, y, z) are matched using least squares. In the perfect world, one would have by Eq. (1)
Random errors have an influence, Eq. (2) is inconsistent.
Consider that by Eq. (4)
Impulse response coefficients (::left{{g}_{k}left(theta:right)right})based on Eq. (5)
and suppose that the impulse response, which we represent by {gˆk}, has yielded the first N coefficients {gk(θ)} that we have measured by Eq. (6)
By minimising the error criterion, we can derive an estimate of θ by Eq. (7)
The following can be used to accomplish the minimising in place of doing it directly. Take note that by revising by Eq. (8)
and enlarging its polynomials, we are able to express the relationship between each coefficient as Eq. (9)
An overview of user interactions inside the blockchain network at each university is presented in Fig. 3. We assume all user devices in this proposed architecture have limited power, memory, and computing capabilities to communicate with the blockchain. Unregistered or unverified user devices cannot authenticate and, as a result, are prohibited from communicating with authorised devices within or outside of the same university. Following this process decreases the likelihood of a malicious device connecting with a legitimate device. Authorised university staff members have public access to all network records and can validate the academic records of any linked university’s students in the blockchain. For example, because their academic information is accessible to all relevant university officials regardless of location, students can enrol in courses, and instructors can apply to teach courses at other connected institutions within the network via the blockchain.
Additionally, the system lets employers or other educational institutions confirm the validity and integrity of the certificates that graduating students receive by allowing them to be issued on the blockchain. Records are made for these transactions in order to confirm the execution of the activities and validate the transactions. Subsequently, the learning transaction records are dispersed throughout the blockchain network to offer distributed authentication and authorisation to users and their registered devices.
User interactions at each university within the blockchain network.
The blockchain-based access control for student academic records is intended to give students authority over who can view their personal health data while offering an effective and safe way to manage their academic records. This architecture creates a decentralised system that guarantees data privacy and protection by utilising the Ethereum blockchain and solidity intelligent contract language. Smart contracts are used in this process to provide and cancel access authorisation. Smart contracts allow for coding interaction rules between entities, which are then automatically carried out when triggered. For the desired task, the contract has three primary mappings: authorizedUsers(), accessPermissions (), and record(). An authorised user list for access to student educational records can be found in the authorizedUsers() mapping. A list of students’ educational records that every user is permitted to access is contained in accessPermission() mapping. The record() mapping, as illustrated in Fig. 3, has a list of students’ academic records hashed on the blockchain. In contrast, contract owners have an even workload to validate each transaction that subjects seek.
The address of student, hash of their academic record, a timestamp, description of transactions are all stored in Record struct. Following is a description of the mappings that were used to implement the smart contract:
authorizedUsers(): an address to boolean mapping. List of authorised users is stored there. This function verifies the presence of the student record. Function notifies other users on network that a user’s access to a student’s educational record is cancelled if record is not present. In event that record is found, a new item is made in authorised user’s mapping with authorised user’s address and a Boolean value designating whether or not user given access to record.
accessPermissions(): a mapping from addresses to booleans via a hierarchical mapping of record hashes. Every user’s and record’s access permissions are stored there. The allowAccess() function sets levels of access authorization for authorised user to true and verifies that student using function is owner of student’s educational record. RevokeAccess() can be triggered if user access permission is set to false.
record(): uses addrecord, updaterecord, and getrecord, to perform new addition, update, obtain details. A student runs the addRecord() function to add a new student’s educational record to blockchain. If there isn’t already a record of the students, the method’s goal is to make one. UpdateRecord() function modifies record’s description and verifies that students calling it are ones who hold educational record. An authorised user must use getRecord() with their address and hash of student’s educational record in order to obtain the student’s record. The getRecord function in Fig. 4 retrieves the student’s educational record if access is allowed to the authorised user after verifying that access has been permitted.
Accessing records utilizing smart contracts.
In order to verify blocks and update blocks in the contract, parallel execution mode is used. Because the block formation and verification processes are carried out in a concurrent execution mode, the network’s contract execution performance also reduces time consumption. Every contract in the block is carried out sequentially by the miner. Control can move from one contract’s code to another contract’s code and back again when one contract calls upon the features of another contract. Concurrent transaction validation is another option. Miners’ proposed transactions may be re-executed by validators in a different order, producing an unanticipated result in block rejection. The two measures from Fig. 5 that are fed into fuzzy logic system are based on semantic similarity between student records and two opinion texts. Tweet’s class is its output. As previously said, defining inputs and outputs—definition of linguistic variables in input and output for our proposed FLS—is the first stage in an FLS.
Fuzzy based malicious user detection.
In this instance, we define two input variables—positivity and negativity of educational data users—and one output variable—class of tweets—because we wish to categorise users of educational data into three classes: positive, negative, neutral. Any variable in an FLS, whether input or output, is called a linguistic variable. Linguistic terms, or fuzzy sets, are the values that any linguistic variable can have. Any variable in an FLS, whether input or output, is called a linguistic variable. Linguistic terms, or fuzzy sets, are the values that any linguistic variable can have. Determining crisp values of inputs that start our method is the next stage in our FLS after we have described linguistic variables as well as their linguistic words in input and output. 50% of the dataset was utilised to train the model, which is then used to map words onto their corresponding vector representations. Softmax probability is computed for each word to determine high-dimensional vectors for each word. Dimension of vector is associated with the quantity of neurons in buried layer. Vector dimension of each word starts at 100. To ensure that each sentence’s length is consistent across the dataset, zero vectors pad it. After that, a sentence vector X = {w1, w2,…,wi,…, w|x|} is constructed for each review x, X ∈ Rd×|x|, where wi denotes word embedding at associated position i in a sentence. The convolutional neural network is then fed X.
Convolution Layer: This layer encompasses each sentence with a sliding window of length h that has a set of m filters applied to it. A feature ci is produced when these filters are applied to each window of words that could possibly exist in the phrase. Every filter has a unique bias of its own. Several feature maps are produced by these m filters operating in concurrently.
Global Max Pooling Layer: The local optimal features and feature map produced by convolution layer are sampled by pooling layer. By combining the data, this layer lessens the representation.
Fully Connected Layer: The Eq. (3), where α is rectified linear unit (ReLU) activation function, W ∈ Rm×m is weight matrix, b ∈ Rm is bias, Cpool is feature map matrix produced by pooling layer, is used by fully connected layer to compute transformation.
Sentence embedding for every review is represented by output vector of this layer. Ultimately, a completely connected softmax layer receives output from preceding layer. The class K with the highest probability is returned. Since error probability between network prediction as well as actual output label is measured in three different classes, “categorical_crossentropy” loss function is employed in the softmax layer. After softmax layer gives classification result, back-propagation method updates the model parameters based on the training data’s actual classification label. Ultimately, three labels with values are assigned to each sentence, one of which corresponds to the actual label. As an illustration, “positive” is equal to [0, 0, 1], “negative” to [1, 0, 0], and “neutral” to [0, 1, 0].
We first explain a generic design for the network’s essential component, convolutional sequence prediction, before specifying the network’s structure⁷. Assume we have a set of malicious userdata {x0,., xT } and we utilise these to forecast the malicious userdata {y0,., yT } at the following period time. To using the current observed data {x0,., xt − 1} as inputs in order to forecast the outputs yt for a certain time t. The following mapping can be defined by the function f: XT → Y T, which represents a sequence modelling network^25,53.
To forecast the malicious user yt at time t, we have the function f if it meets causal constraint that yt depends only on {x0,…, xt − 1} rather than any “future”’ inputs {xt + 1,…, xT } by Eq. (10)
Finding a network f that can minimise expected loss between actual data and prediction, or L (yˆt, f(x0,., xt − 1)), is the aim of learning the sequence modelling configuration. Suggested deep learning system is feed-forward convolutional temporal neural network (FCTNN). It was initially created for action segmentation and detection²⁶, and this is where our FCTNN model receives inspiration and adaptations from. The FCTNN stands out from other neural networks that are currently in use for short-term harmful user forecasting. This trait logically matches the sequence prediction mentioned above^14,27. At time t, the causal convolutions essentially function as a filter that can only view inputs that are received no later than t. This prevents knowledge from leaking from the past to the present. This enables our framework to process input of any length for a data sequence²⁸. It requires extensive networks with a lengthy effective history. This is certain to result in a convoluted network architecture and significant processing overhead²⁹. Alternatively, the suggested FCTNN architecture incorporates residual layers and dilated convolutions. Dilated convolutions in particular allow for an increasingly wide receptive field by Eq. (11)
where t − d · i denotes direction of past, d is dilation factor, and k is the filter size. Every two consecutive filter taps, there is a fixed step that represents the dilation factor. In reality, a regular convolution is a dilated convolution with a dilation factor of d = 1. Dilation factor modifies the TCN’s receptive field. Our methodology ensures an extraordinarily broad effective history by adjusting d exponentially with network depth. As a result, the receptive field can expand due to the increase in dilatation³⁰. As a result, a larger range of inputs are represented in the output at the top level. Keep in mind that you may also change the filter size k to expand the FCTNN’s receptive field.
We employed the rectified linear unit (ReLU) for each of the two weight layers in the FCTNN. Furthermore, a spatial dropout for regularisation is included following the final weight layer. Formally, the residual block is defined as follows in this paper by Eq. (12)
Y represents the layer’s output vector in this instance. Two layers are represented by the formula F = W2σ(W1x) + e, where σ stands for ReLU and e for bias. The FCTNN architecture for our framework is built based on the original FCTNN setup, as described in this paper. A set of blocks, each containing a succession of L convolutional layers, composites it. Dilated convolutions, which are connected to a non-linear activation f and a dilation factor d, combine each layer (.). In addition, each dilated convolution has a residual link added to it in order to integrate the layer’s input with the convolution result. Assume that S (i, j) ∈ R Fw×T represents activations for ith layer and jth block. Observe that each layer I has the same number of filters (Fw). We can use processed training data to train TCN method once it has been constructed. The central server will receive the final TCN with optimised structure, which will identify malevolent users.
In private blockchains, smart contracts offer a potent means of implementing role-based access control. They are a crucial part of any reliable and safe blockchain system because they offer transparency, automation^31,32, and flexibility in the management of access control regulations. Smart contracts can be utilised on a private blockchain to implement RBAC by specifying the roles and permissions of individual users³³. These roles can be used to limit access to specific system data or operations^34,54. It is possible to construct the smart contract so that only users who are assigned proper roles can access or carry out particular operations on the contract.
Free travel (Exploration): The global search is carried out by the SRSO using the Sailed Fish Optimizer (SFO) approach, which is based on the elite method employed in the swordfish algorithm. The following is an expression for the position updating formula by Eq. (13)
And the ith remora’s candidate position is represented by Vi(t + 1). The best position as of right now is Xbest(t). Remora’s random position is denoted by Xrand(t). Iteration number t is what we’re talking about. A random number between 0 and 1 is called a rand. Furthermore, remora has the ability to switch hosts based on its experiences. In this instance, a fresh candidate position by Eq. (14)
SFO Strategy: It is evident that the remora will move with the sailfish once it has adsorbed on it. With the use of the SFO algorithm’s elite approach, the formula is enhanced and yields the following formula by Eq. (15)
where XBest t is current best position, Xrand t is current random position of Remora, t is iteration number.
Experience Attack: After being adsorbed on the host, the remora will explore a small portion of the host by using the locations of previous and present generations of remora to determine whether host needs to be replaced. This method is similar to experience-building process. Formula for a mathematical calculation is shown in Eq. (16)
where Xatt represents the remora’s tentative movement. As an experience, Xpre might be thought of as the position of the preceding generation of Remora. Lastly, a random number with a normal distribution between 0 and 1 is called rand. The remora uses Eq. (17) to determine whether to switch hosts after a brief range of motion and defines the decision method.
H(i), among them, has an initial value of 0 or 1 and indicates the host absorbed by remora. Whale is absorbed if H(i) equals 1, sailfish is adsorbed if H(i) equals 0. Additionally, round is a rounded function, and the fitness values of Xi t and Xatt are, respectively, f(Xi t) and f(Xatt).
Every experiment in this research was carried out using MATLAB R2021a on a Windows 11 PC with an Intel (R) Core (TM) i7-11700 CPU operating at 2.50 GHz and 16 GB of RAM.
CERT Dataset: We utilised the “CERT Insider Threat Tools” dataset (Carnegie Mellon’s Software Engineering Institute, Pittsburgh, PA, USA) since it is quite difficult to obtain actual business system logs. CERT dataset is an intentionally manufactured dataset utilised to validate insider-threat detection systems; it is not real-world^35,36 corporate data. Employee computer usage logs and specific organisational data like employee departments and roles are included in the CERT dataset. Every table has data about each user’s activities, timestamps, and ID. The CERT dataset is available in six significant versions (R1 through R6), with R6.1 and R6.2 being the most recent releases. Depending on the version of the dataset, there are differences in the kinds of usage data³⁷, the number of variables, the number of employees, and the quantity of malicious insider activities. The largest and most recent dataset, R6.2, was used for this investigation—only five of the 4000 users in this version of the dataset engaged in harmful behaviour. Table 1 describes the logon activity table.
User Activity: This research aims to monitor student behaviour throughout the learning process to assess elements that might affect a student’s academic achievement. There are 150 student records with 11 attributes in the gathered data set. Three primary categories classify the features: (1) Gender and nationality are examples of demographic features. (2) Academic background elements, including section, grade, and stage. (3) Behavioural elements include raising one’s hand during class, accessing resources, participating in discussion groups, and paying attention to messages and announcements. One of the key areas of research in educational psychology is student involvement. “The quality and quantity of students’ psychological, cognitive, emotional, and behavioural reactions to the learning process as well as to in-class/out-of-class academic and social activities to achieve successful learning outcomes” is the definition given to student engagement. Attributes and features of the dataset are displayed in Table 2, along with a description. A behavioural factor is a new feature category visible in the table. These qualities have to do with the experiences that students have and how they behave when they are in school. We use a few preprocessing methods to improve the data set’s quality after the data collection activity. Data preprocessing, which encompasses data transformation, data reduction, data cleansing, and feature selection, is a crucial phase in knowledge discovery.
We looked into the roles of the 73 anomalous events, as indicated in Table 3, to determine traits of malevolent insiders. It was discovered that the three roles of “salesman,” “information technology (IT) administrator,” and “electrical engineer” account for almost 90% of the most aberrant actions. Constructing an effective detection model without any abnormal instances in a role is challenging. However, it is also impossible to validate the performance of the produced method in case of roles with fewer than three abnormal instances. The frequency of normal and abnormal instances in 3 roles are shown in Table 3.
Table 4 and Figure6(a)-(d) compare a classifier based on various feature classes of input student educational data. Here, the feature classes³⁹ analyzed are demographic features, academic background features, behavioral factors in prediction accuracy, MAP, F-1 score, And latency. The proposed technique achieved a prediction accuracy of 98%, a mean average precision (MAP) of 95%, an F1 score of 97%, And a latency of 96% for demographic features. For academic background features, it attained a prediction accuracy of 94%, a MAP of 91%, an F1 score of 95%, And a latency of 94%. Regarding behavioural factors, the proposed technique achieved a prediction accuracy of 89%, a MAP of 88%, an F1 score of 93%, And a latency of 92%.
In contrast, existing classifiers—CNN, KNN, SVM, And random Forest—obtained a prediction accuracy of 68%, a MAP of 72%, an F1 score of 63%, And a latency of 71% for demographic features. For academic background features, these classifiers achieved a prediction accuracy of 73%, a MAP of 76%, an F1 score of 67%, And a latency of 74%. For behavioural factors, they attained a prediction accuracy of 75%, a MAP of 73%, an F1 score of 71%, And a latency of 77%.
Comparative analysis based on feature class for various classifiers in terms of (a) Prediction accuracy, (b) MAP, (c) F-1 score, (d) Latency
Comparative analysis based on blockchain security analysis in terms of (a) QoS, (b) Contract execution time, (c) Throughput
As shown in Table 5, the proposed B-FCTNN_SRSO achieved a QoS of 97%, a precision of 94%, and a throughput of 96% based on blockchain security analysis, as illustrated in Fig. 7(a)-(c). In comparison, the existing KNN technique attained a QoS of 92%, a contract execution time of 89%, and a throughput of 85%, while SVM achieved a QoS of 95%, a contract execution time of 93%, and a throughput of 89%.
The Performance comparison of Training, Testing, Validation Accuracy and Performance Stability with various models.
Table 6 compares various machine learning models based on their performance stability across dataset splits and their training, validation, and testing accuracy. Overfitting^40,41 occurs when a model performs well on training data but poorly on unseen data. As shown in Fig. 8, CNN has the lowest testing accuracy at 74%. KNN improves upon this with a validation accuracy of 82% and a testing accuracy of 79%, though its stability remains 82%, indicating minor discrepancies across dataset splits⁴². SVM further enhances testing accuracy to 85%, but inconsistencies between training and validation suggest challenges with generalization⁴³. Random Forest performs well, achieving an 87% testing accuracy and 89% stability. The proposed B-FCTNN_SRSO model outperforms all others, achieving the highest stability (96%), training accuracy (98%), validation accuracy (96%), and testing accuracy (95%).
Table 7 highlights the proposed B-FCTNN_SRSO model, comparing various techniques based on multiple performance metrics. False positive rates (FPR) and false negative rates (FNR) are key indicators of detecting malicious users effectively. As shown in Fig. 9, the proposed model achieves a significantly lower FPR (2%) and FNR (5%) compared to CNN (FPR: 10%, FNR: 15%), demonstrating superior accuracy in threat detection. Regarding memory efficiency, the B-FCTNN_SRSO model outperforms CNN (550 MB) and KNN (520 MB), utilizing only 430 MB of memory. Its drastically reduced training time of 1.2 s makes it highly suitable for real-time applications^43,44. Additionally, it enhances blockchain system efficiency by minimizing storage overhead to just 0.8 MB and reducing validation time to 85 ms. The proposed model maintains lightweight data⁴⁵ handling compared to CNN (1.8 MB) due to its lower storage requirements. Furthermore, the B-FCTNN_SRSO model offers the highest adversarial resistance (92%), making it more resilient to cyber threats. This upgrade significantly enhances the security of blockchain-based document management, surpassing CNN (72%) and KNN (78%). Overall, the proposed model delivers superior accuracy, efficiency⁴⁶, speed, and security, making it the most effective solution for managing educational documents electronically.
The comparison of proposed method thru advanced parameters against state of art methods.
Table 8 compares different methods based on computational complexity, security, scalability, and blockchain integration efficiency. The model GANs and RNNs have high computational complexity⁴⁷ and limited blockchain integration, making them less efficient for secure document management. Hybrid Federated Learning provides high security and scalability but requires significant computational resources⁴⁸. In contrast, the proposed B-FCTNN_SRSO offers low computational complexity, high security through blockchain authentication, and excellent scalability⁴⁹. Its smart contract-based access control ensures efficient blockchain integration, making it more suitable for secure and scalable electronic document management⁵⁰.
Table 9 compares PSO, Genetic Algorithm, and the SRSO proposed regarding convergence speed, computational complexity, optimization efficiency, scalability⁵⁰, and access control security. The PSO offers average convergence speed but consumes high computational power and optimized parameters, thus making it less efficient for access control⁵¹. Genetic Algorithms offer high optimization efficiency but slow convergence speed and high computational complexity, thus prone to premature convergence. SRSO, however, is the fastest, offers low computational complexity, and is highly scalable and secure⁵². Its dynamic adaptation of role-based access control makes it the most efficient solution for secure blockchain applications.
The Security Analysis against the Various Models.
Table 10 compares various algorithms regarding blockchain security risks, including Sybil attack resistance, privacy protection, scalability, and overall security efficiency.
PSO and Genetic Algorithms perform poorly securing blockchain-based systems, exhibiting only moderate smart contract security and weak Sybil attack resistance (50–65%). Additionally, privacy protection remains inadequate (50–55%), posing a significant risk of data leakage. In contrast, Federated Learning offers strong privacy protection (90%) and scalability (85%), ensuring better data security. However, its weak smart contract security (50%) leaves it vulnerable to contract exploits, reducing its reliability. As illustrated in Fig. 10, the proposed SRSO-Based Model outperforms all other approaches, achieving 95% Sybil attack resistance, 90% smart contract security, and 95% privacy protection. Its high scalability (92%) ensures effective role-based access control in blockchain-based educational document management. Regarding enhancing security and efficiency in blockchain applications, SRSO stands out as the most scalable and secure solution.
Comparative Analysis Against the existing Blockchain based Credential Verification Systems.
As per Fig. 11, a comparative analysis is performed against the existing Blockchain based credential verification systems. Table 11 compares various blockchain-based credential verification systems based on efficiency, security, speed, attack resistance, accuracy, and smart contract optimization. The Hyperledger Fabric-Based System is less efficient for large-scale applications due to its slower verification time (120 ms) and lower accuracy (85%). The Ethereum Smart Contract System improves verification speed to 110 ms and accuracy to 88%, but its attack resistance remains relatively low at 78%. While the Decentralized Identity (DID) System further enhances security (87%) and attack resilience (82%), its verification speed remains moderate at 100 ms.As illustrated in Fig. 11, the proposed B-FCTNN_SRSO technique outperforms all competing methods, achieving higher accuracy (98%), faster verification speed (85 ms), improved scalability (95%), and superior security efficiency (96%).
A system that uses machine learning algorithms to analyse and improve the management of educational documents, such as transcripts, student records, and certificates, stored safely on a blockchain network. This system offers features like automated verification, personalised learning insights, and fraud detection, all while preserving data integrity and transparency throughout the educational ecosystem.
Scalability: It cannot be easy to handle massive amounts of educational data on a blockchain, mainly when dealing with intricate learning algorithms.
Adoption Barriers: It might be necessary to make considerable adjustments to infrastructure and protocols to incorporate blockchain technology into current educational systems.
Privacy Issues: The advantages of blockchain transparency must be carefully weighed against data privacy.
The novel method for managing educational documents was proposed through a blockchain-based fuzzy feed-forward convolutional temporal neural network to detect malicious users. This instance uses NLP analysis to document word weight indexing during training. The role-based access control combined with simulated remora swarm optimization enforces document management access control. The suggested architecture aims to identify rogue users by authenticating users and logging access requests on the blockchain. This may be the first study to integrate student behavior with academic success. The high accuracy results validate the newly acquired information from categorization approaches, which finds that learners’ activities played a significant role in the learning process. Based on user behavior modelling and anomaly detection methods, we suggested an insider-threat detection framework. Individual users’ diverse behaviors are converted into a structured dataset throughout the user behavior modelling process, where each row corresponds to an instance, and every column corresponds to input variables for anomaly detection methods. We identified future research directions due to the current study’s limitations, even though the proposed framework was empirically verified. Our insider-threat detection model was constructed using a specific time unit, such as a day. In another way, this method can identify harmful activities based on batch process but cannot instantly identify them. Therefore, creating an online stream data-processing sequence-based insider-threat detection method could be worthwhile. In order to obtain more accurate results, our future work will involve applying data mining algorithms on an expanded data set with additional distinguishing qualities.
The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.
Pandiaraja, P., Boopesh, K. B., Deepthi, T., Laksmi Priya, M. & Noodhana, R. An analysis of document summarization for educational data classification using NLP with machine learning techniques. In International Conference on Computing in Engineering & Technology (pp. 127–143). Singapore: Springer Nature Singapore. (2022), February.
Yağcı, M. Educational data mining: prediction of students’ academic performance using machine learning algorithms. Smart Learn. Environ. 9 (1), 11 (2022).
Article  Google Scholar 
Pandey, M. et al. AI-based integrated approach for the development of intelligent document management system (IDMS). Procedia Comput. Sci. 230, 725–736 (2023).
Article  Google Scholar 
Omurca, S. I. et al. A document image classification system fusing deep and machine learning models. Appl. Intell. 53 (12), 15295–15310 (2023).
Article  Google Scholar 
Zhou, J. & Huang, T. Application of machine learning algorithm in electronic book database management system. SN Appl. Sci. 5 (11), 287 (2023).
Article  Google Scholar 
Alam, A. A digital game-based learning approach for effective curriculum transaction for teaching-learning of artificial intelligence and machine learning. In 2022 International Conference on Sustainable Computing and Data Communication Systems (ICSCDS) (pp. 69–74). IEEE. (2022), April.
Zanbouri, K. et al. A GSO-based multi‐objective technique for performance optimization of blockchain‐based industrial internet of things. Int. J. Commun Syst, 37(15), e5886. (2024).
Dewangan, N. K. & Chandrakar, P. Implementing blockchain and deep learning in the development of an educational digital twin. Soft. Comput., 1–18. (2023).
Wu, X. Y. Exploring the effects of digital technology on deep learning: a meta-analysis. Educ. Inform. Technol. 29 (1), 425–458 (2024).
Article  Google Scholar 
Kaddoura, S., Popescu, D. E. & Hemanth, J. D. A systematic review on machine learning models for online learning and examination systems. PeerJ Comput. Sci. 8, e986 (2022).
Article  PubMed  PubMed Central  Google Scholar 
Razzaq, S. et al. DeepClassRooms: a deep learning based digital twin framework for on-campus class rooms. Neural Comput. Appl., 1–10. (2023).
Zhang, Q. Secure preschool education using machine learning and metaverse technologies. Appl. Artif. Intell. 37 (1), 2222496 (2023).
Article  Google Scholar 
Rajendran, R. et al. An exploratory analysis of machine learning adaptability in big data analytics environments: A data aggregation in the age of big data and the internet of things. In 2022 2nd International Conference on Innovative Practices in Technology and Management (ICIPTM) (Vol. 2, pp. 32–36). IEEE. (2022), February.
Amiri, Z., Heidari, A. & Navimipour, N. J. Comprehensive survey of artificial intelligence techniques and strategies for climate change mitigation. Energy, 132827. (2024).
Fahd, K., Venkatraman, S., Miah, S. J. & Ahmed, K. Application of machine learning in higher education to assess student academic performance, at-risk, and attrition: A meta-analysis of literature. Educ. Inform. Technol., 1–33. (2022).
Shi, Y., Sun, F., Zuo, H. & Peng, F. Analysis of learning behavior characteristics and prediction of learning effect for improving college students’ information literacy based on machine learning. IEEE Access. (2023).
Kumar, D. N. & Kavuri, R. Hybrid digital certificate management system with QR code and IoT integrated on hyperledger fabric blockchain. Int. J. Maritime Eng. 1 (1), 635–646 (2024).
Google Scholar 
Xie, R. et al. June., Ethereum-Blockchain-Based Technology of Decentralized Smart Contract Certificate System, in IEEE Internet of Things Magazine, vol. 3, no. 2, pp. 44–50, (2020). https://doi.org/10.1109/IOTM.0001.1900094
Deng, H. et al. FutureDID: A fully decentralized identity system with Multi-Party verification. IEEE Trans. Comput. 73 (8), 2051–2065. https://doi.org/10.1109/TC.2024.3398509 (Aug. 2024).
Vakili, A. et al. A new service composition method in the cloud‐based internet of things environment using a grey Wolf optimization algorithm and mapreduce framework. Concurrency Computation: Pract. Experience. 36 (16), e8091 (2024).
Article  Google Scholar 
Heidari, A., Shishehlou, H., Darbandi, M., Navimipour, N. J. & Yalcin, S. A reliable method for data aggregation on the industrial internet of things using a hybrid optimization algorithm and density correlation degree. Cluster Comput. 27 (6), 7521–7539 (2024).
Article  Google Scholar 
Heidari, A., Navimipour, N. J. & Unal, M. A secure intrusion detection platform using blockchain and radial basis function neural networks for internet of drones. IEEE Internet Things J. 10 (10), 8445–8454 (2023).
Article  Google Scholar 
Heidari, A., Amiri, Z., Jamali, M. A. J. & Jafari, N. Assessment of reliability and availability of wireless sensor networks in industrial applications by considering permanent faults. Concurrency Computation: Pract. Experience, 36(27), e8252. (2024).
Amiri, Z., Heidari, A., Zavvar, M., Navimipour, N. J. & Esmaeilpour, M. The applications of nature-inspired algorithms in internet of Things‐based healthcare service: A systematic literature review. Trans. Emerg. Telecommunications Technol., 35(6), e4969. (2024).
Heidari, A., Navimipour, N. J., Dag, H., Talebi, S. & Unal, M. A novel blockchain-based deepfake detection method using federated and deep learning models. Cogn. Comput. 16 (3), 1073–1091 (2024).
Article  Google Scholar 
Heidari, A., Navimipour, N. J., Zeadally, S. & Chamola, V. Everything you wanted to know about ChatGPT: components, capabilities, applications, and opportunities. Internet Technol. Lett., 7(6), e530. (2024).
Asadi, M., Jamali, M. A. J., Heidari, A. & Navimipour, N. J. Botnets unveiled: A comprehensive survey on evolving threats and defense strategies. Trans. Emerg. Telecommunications Technol., 35(11), e5056. (2024).
Heidari, A., Jamali, M. A. J. & Navimipour, N. J. Fuzzy logic multicriteria Decision-Making for broadcast storm resolution in vehicular ad hoc networks. Int. J. Commun Syst, 38(5), e6034. (2025).
Toumaj, S., Heidari, A., Shahhosseini, R. & Jafari Navimipour, N. Applications of deep learning in Alzheimer’s disease: A systematic literature review of current trends, methodologies, challenges, innovations, and future directions. Artif. Intell. Rev. 58 (2), 44 (2024).
Article  Google Scholar 
Amiri, Z., Heidari, A., Jafari, N. & Hosseinzadeh, M. Deep study on autonomous learning techniques for complex pattern recognition in interconnected information systems. Comput. Sci. Rev. 54, 100666 (2024).
Article  Google Scholar 
Heidari, A., Jamali, M. A. J. & Navimipour, N. J. An innovative performance assessment method for increasing the efficiency of AODV routing protocol in VANETs through colored timed petri Nets. Concurrency Computation: Pract. Experience. 37 (3), e8349 (2025).
Article  Google Scholar 
Heidari, A., Jabraeil Jamali, M. A., Navimipour, J., Akbarpour, S. & N., & Deep Q-learning technique for offloading offline/online computation in blockchain-enabled green IoT-edge scenarios. Appl. Sci. 12 (16), 8232 (2022).
Article  CAS  Google Scholar 
Kirubasri, G. et al. A recent survey on 6G vehicular technology, applications and challenges. In 2021 9th International conference on reliability, infocom technologies and optimization (trends and future directions)(ICRITO) (pp. 1–5). IEEE. (2021), September.
Pandey, B. K. & Pandey, D. Parametric optimization and prediction of enhanced thermoelectric performance in co-doped CaMnO3 using response surface methodology and neural network. J. Mater. Sci.: Mater. Electron. 34 (21), 1589 (2023).
CAS  Google Scholar 
Anand, R., Lakshmi, S. V., Pandey, D. & Pandey, B. K. An enhanced ResNet-50 deep learning model for arrhythmia detection using electrocardiogram biomedical indicators. Evol. Syst. 15 (1), 83–97 (2024).
Article  Google Scholar 
Pandey, B. K., Pandey, D. & Sahani, S. K. Autopilot control unmanned aerial vehicle system for sewage defect detection using deep learning. Eng. Rep., 7(1), e12852. (2025).
Kumar Pandey, B. et al. Encryption and steganography-based text extraction in IoT using the EWCTS optimizer. Imaging Sci. J. 69 (1–4), 38–56 (2021).
Article  Google Scholar 
Raja, D. S. S., Kumar, D. R., Santhiyakumari, N., Kumarganesh, S., Sagayam, K. M.,Thiyaneswaran, B., … Pandey, D. (2024). A compact dual-feed wide-band slotted antenna for future wireless applications. Analog Integrated Circuits and Signal Processing,118(2), 291–305.
Dhanasekar, S., Sagayam, M., Pandey, K., Pandey, D. & B. K., & Refractive index sensing using metamaterial absorbing augmentation in elliptical graphene arrays. Plasmonics 19 (5), 2273–2283 (2024).
Article  CAS  Google Scholar 
Maheshwari, R. U., Kumarganesh, S., KVM, S., Gopalakrishnan, A., Selvi, K., Paulchamy,B., … Pandey, D. (2024). Advanced plasmonic resonance-enhanced biosensor for comprehensive real-time detection and analysis of deepfake content. Plasmonics, 1–18.
Pandey, D. & Pandey, B. K. An efficient deep neural network with adaptive galactic swarm optimization for complex image text extraction. In Process Mining Techniques for Pattern Recognition (121–137). CRC. (2022).
Pandey, B. K., Pandey, D. & Lelisho, M. E. Face mask identification with enhanced cuckoo optimization and deep learning-based faster regional neural network. Sci. Rep. 14 (1), 29719 (2024).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Du John, H. V., Jose, T., Sagayam, S. A. G., Pandey, K. M., Pandey, D. & B. K., & Enhancing absorption in a metamaterial absorber-based solar cell structure through anti-reflection layer integration. Silicon 16 (9), 3811–3821 (2024).
Article  CAS  Google Scholar 
Abdullahi, M., Ahmad, A., Pandey, B. K. & Pandey, D. Digital currency adoption: a comparative analysis of global trends and the Nigerian Enaira digital currency. SN Comput. Sci. 5 (5), 602 (2024).
Article  Google Scholar 
Vinodhini, V. et al. IoT-based early forest fire detection using MLP and AROC method. Int. J. Global Warming. 27 (1), 55–70 (2022).
Article  Google Scholar 
Sharma, M., Pandey, D., Palta, P. & Pandey, B. K. Design and power dissipation consideration of PFAL CMOS V/S conventional CMOS based 2: 1 multiplexer and full adder. Silicon 14 (8), 4401–4410 (2022).
Article  CAS  Google Scholar 
Revathi, T. K. et al. An intelligent model for coronary heart disease diagnosis. In Networking Technologies in Smart Healthcare (309–327). CRC. (2022).
Pandey, B. K. & Pandey, D. Securing healthcare medical image information using advance morphological component analysis, information hiding systems, and hybrid convolutional neural networks on IoMT. Comput. Biol. Med. 185, 109499 (2025).
Article  PubMed  Google Scholar 
Khadka, M. et al. Spectroscopic, quantum chemical, and topological calculations of the phenylephrine molecule using density functional theory. Sci. Rep. 15 (1), 208 (2025).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Maheshwari, R. U., Pandey, A. R. J., Pandey, D. & B. K., & Innovative Quantum PlasmoVision-Based Imaging for Real-Time Deepfake Detection1–17 (Plasmonics, 2025).
Praveenkumar, T., Anthoniraj, S., Kumarganesh, S., Somaskandan, M., Martin Sagayam,K., Pandey, B. K., … Sahani, S. K. (2025). Enhanced Circuit Board Analysis: Infrared Image Segmentation Utilizing Markov Random Field (MRF) and Level Set Techniques. Engineering Reports, 7(3), e70029.
Ahathiyan, G. S. et al. Design and simulation of a highly efficient eco-friendly, non-toxic perovskite solar cell. Discover Nano. 20 (1), 32 (2025).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Sheela, M. S., Kumarganesh, S., Pandey, B. K. & Lelisho, M. E. Integration of silver nanostructures in wireless sensor networks for enhanced biochemical sensing. Discover Nano. 20 (1), 7 (2025).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Pandey, D., Sharma, M., Talwar, R. & Pandey, B. K. A compact quintuple band miniaturized elliptical planar monopole antenna for 5G/6G wireless systems. Analog Integr. Circuits Signal Process. 122 (3), 1–16 (2025).
Article  Google Scholar 
Satheesh, N., Gopisankar, N., Kumarganesh, S., Anthoniraj, S., Saravanakumar, S.,Sagayam, K. M., … Pandey, D. (2025). Advanced AI-driven emergency response systems for enhanced vehicle and human safety. Iran Journal of Computer Science, 1–16.
Download references
The authors would like to express gratitude to Department of Statistics, College of Natural and Computational Science, Mizan-Tepi University, Tepi, Ethiopia.
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Department of Computer Science and Engineering, School of Computing, Kalasalingam Academy of Research and Engineering, Srivilliputtur, India
P. Chinnasamy
Department of Data Science and Business Systems, School of Computing, SRMIST, Kattankulathur, Chennai, India
B. Subashini
Faculty of Information and Communication Technology, Universiti Tunku Abdul Rahman, Kampar, Perak, Malaysia
Ramesh Kumar Ayyasamy
Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad, India
Ajmeera Kiran
Department of Information Technology, College of Technology, Govind Ballabh Pant University of Agriculture and Technology Pantnagar, Uttrarakhand, India
Binay Kumar Pandey
Department of Technical Education (Government of Uttar Pradesh), Kanpur, Uttar Pradesh, India
Digvijay Pandey
Department of Statistics, College of Natural and Computational Science, Mizan-Tepi University, Tepi, Ethiopia
Mesfin Esayas Lelisho
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
P. Chinnasamy and Subashini. B perform experiment, wrote manuscript. Ramesh Kumar Ayyasamy and Ajmeera Kiran draw figure, wrote manuscript. Binay Kumar Pandey and Digvijay Pandey develop methodology, wrote manuscript. Mesfin Esayas Lelisho conceptualize the work, wrote manuscript.
Correspondence to Mesfin Esayas Lelisho.
The authors declare that they have ‘no known conflict of interests or personal relationships’ that could have appeared to influence the work reported in this paper.
Not Applicable (as the results of studies do not involve any human or animal).
Not Applicable (as the results of studies do not involve any human or animal).
Not Applicable (as the results of studies does not involve any human or animal).
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Reprints and permissions
Chinnasamy, P., Subashini, B., Ayyasamy, R.K. et al. Blockchain based electronic educational document management with role-based access control using machine learning model. Sci Rep 15, 18828 (2025). https://doi.org/10.1038/s41598-025-99683-5
Download citation
Received: 07 December 2024
Accepted: 22 April 2025
Published: 29 May 2025
DOI: https://doi.org/10.1038/s41598-025-99683-5
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2025 Springer Nature Limited
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

source

The Trump administration’s funding freeze affecting several prominent international education grant programmes in the US continues. Over 10,000 students and professionals participating in international exchanges – some American, some from other countries – have had their funding withdrawn. They have been given no indication of when – or if – funding of their programmes will be reinstated.
The US government paused funding of all programmes under the State Department’s Bureau of Educational and Cultural Affairs (ECA) on 13 February 2025 for 15 days. Instead of ending the pause as expected on 27 February, it kept the freeze in place. Affected programmes include:
Of the students currently affected, more than 3,500 are abroad. Another 7,400 are international students in the US. Funding has also been paused for students planning to go on exchanges in the next six months.
Many students have been left wondering how they will pay their rent, and some are resorting to food banks for meals.
In addition, the ECA funding normally pays for the international education professionals employed by the scholarship and exchange programmes. Those staff are understandably worried about losing their jobs.
This week, international education peak bodies announced a joint campaign aimed at persuading Congress to intervene to stop the funding freeze.
In a statement, NAFSA executive director and CEO Dr. Fanta Aw pointed out that the affected programmes have been authorised and appropriated funds by the US Congress, and articulated what is at stake with the funding pause:
“The freeze on State Department grant programs threatens the survival of study abroad and international exchange programs that are essential to U.S. economic and national security. Halting inbound and outbound exchanges shuts the United States off from a vital flow of ideas, innovation, and global understanding and influence, creating a vacuum that could easily be filled by competing nations. We urge Congress to use its authority to intervene. Restoring this funding immediately is absolutely in the country’s national interest.”
Mark Overmann, executive director of the Alliance for International Exchange, emphasised that the funding freeze is not only hurting international students, but also Americans in exchange programmes:
“Paralyzing ECA-funded exchange programs endangers the health, safety, and future of the more than 12,500 Americans who are either abroad right now or soon will be and damages our relationships with current and future leaders from around the world. The many U.S. organizations that support these programs and its participants are now in a dire financial position, putting thousands of American jobs and livelihoods at risk. Approximately 90 percent of the State Department exchanges budget is spent on Americans or in America. ECA exchange programs absolutely fulfill Secretary of State Marco Rubio’s goal of making ‘America safer, stronger, and more prosperous.’ Suspending them would only have the opposite effect.”
Melissa Torres, president and CEO of the Forum on Education Abroad, said:
“Study abroad programs like the Gilman and Fulbright Scholarships provide opportunities for students who might not otherwise have a global education. The Critical Language Scholarship and IDEAS programs build the language skills of U.S. students and broaden their access to destinations where American engagement is particularly important. These experiences help prepare students for a globally connected workforce. Without federal support, the cultural competency of our domestic population and in turn, U.S. global competitiveness, will take a huge step backward.”
Karin Fisher, in a 5 March Latitudes column for the Chronicle of Higher Education, detailed the mood at this week’s Association of International Education Administrators conference in Houston, Texas. Dominant themes were fear and stress. Ms Fisher reports that, “One administrator said his campus was scrambling to find money to pay a Fulbright scholar who was the campus’s sole Chinese-language instructor.”
Another attendee, John Sunnygard, associate provost for global learning and international affairs at Western Kentucky University, expressed his concern for his students who had been awarded Gilman scholarships. The Gilman scholarships are merit-based scholarships that allow students with demonstrated financial need to study or intern abroad. As per Ms Fisher’s reporting, Mr Sunnygard said that many of his Gilman recipients “had already purchased nonrefundable plane tickets and paid program deposits for summertime international study.”
Mr Sunnygard said: “They were promised a scholarship and made a financial commitment. These are kids from rural areas who have never been on a plane. We bought their passports.”
The Chronicle of Higher Education was made privy to a message in which international students were told to leave the US because their funding was “subject to an immediate stop.”
At the conference, Ms Fisher listened to an international graduate student explain that she had received “only a quarter of her regular monthly funding stipend … leaving her unable to pay her rent [and] relying on a food bank.” That student said: “I’ve got zero idea how to pay for my apartment. I’ve got zero idea about what to do with my belongings if I get evicted.”
Outside of the conference, as reported in the Financial Express, Fulbright student Nigora Jabarova posted on LinkedIn that her funding organisation, the Institute of International Education (IIE) had received no warning of the funding freeze. Cut off from funding, she posted:
“With just three months left until my graduation, I now find myself in an extremely difficult position, both professionally and personally. As Fulbright participants, we are not permitted to work outside the program, and the stipend we rely on has been abruptly halted. This has left many of us struggling to cover basic necessities such as rent and food.”
Every year, over 2,000 Indian students receive grants from the Fulbright programme. Vietnam’s VNExpress warns that “Any reduction or delay in funding could limit participation, affecting cultural exchange and academic collaboration between India and the U.S.”
Just last year, former President Biden had made it a priority to forge deeper ties with India through international education.

source

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 15, Article number: 19811 (2025) Cite this article
552 Accesses
1 Altmetric
Metrics details
This study assessed the efficacy of various diagnostic indicators and machine learning (ML) models in predicting childhood myopia. A total of 2,365 children aged 5–12 years were included in the study. The participants were exposed to non-cycloplegic and cycloplegic refraction tests, along with ocular biometric assessments. Cycloplegia was induced using 1% cyclopentolate eye drops, followed by cycloplegic refraction testing. Myopia prevalence was 11.2% (95% confidence interval: 9.9–12.5%). The spherical equivalent (SE) before and after cycloplegia varied with age, significantly differing by 0.5D in children < 10 years (P < 0.05). The most effective single-indicator screening diagnostic methods were axial length/ corneal curvature radius (AL/CCR) and screening myopia, with area under curve (AUC) of 0.919 (95% CI: 0.899 to 0.939) and 0.911 (95% CI: 0.890 to 0.932). In the multi-indicator joint diagnostic model, the best diagnostic model using non-cycloplegic SE, uncorrected distance visual acuity (UCDVA), AL, and age was the Extreme Gradient Boosting model, with an AUC of 0.983 and an accuracy of 0.970. The best diagnostic model using non-cycloplegic SE, AL/CCR, UCDVA, and age was the Random Forest model, with an AUC of 0.981 and an accuracy of 0.975. The AL/CCR demonstrated superior performance in predicting childhood myopia. The ML-based multi-indicator joint diagnostic predictive model enhances the accuracy of childhood myopia diagnosis, screening, and intervention.
Myopia is a major public health concern¹. Its rapidly increasing global prevalence and incidence haves garnered considerable attention from the international community². Its prevalence rose from 28.3% in 2010 to 34% in 2020, an approximate 20% increase from the baseline. By 2050, an estimated 4.758 billion people will be affected, with 938 million at risk of high myopia, representing 49.8% and 9.8% of the global population, respectively³. Myopia adversely affects quality of life, leading to substantial medical expenses and productivity losses amounting to hundreds of billions of dollars annually. Without intervention, these costs will continue to rise⁴.
The earlier the onset of myopia, the higher the risk of developing high myopia and associated ocular complications⁵. Therefore, the early stages of myopia in children are critical for prevention and control. Effective management of myopia in young children can significantly delay its onset.
Traditional large-scale myopia surveys often rely on non-cycloplegic refractive data, which may misrepresent the refractive status of young children and obscure risk factors^1,6. Cycloplegic refraction, though the diagnostic gold standard⁷is costly, invasive, and difficult to administer in children, limiting its use in population-wide studies⁸.
Exploring non-invasive, high-efficacy screening methods for myopia in children is essential. Some experts consider axial length (AL) and corneal curvature radius (CCR) as indicators of the growth and development of children and adolescents^9,10. Their measurements are not affected by the adjustment state and provide objective and accurate data without cycloplegic paralysis. Ocular biometric parameters are important indicators for assessing the far-vision reserve in children and objectively evaluating the development of refraction¹¹.
Traditional linear or logistic regression studies^12,13 have predominantly focused on the screening efficacy of single indicators¹⁴proving unable to handle the nonlinear features of the axial length-to-corneal curvature ratio or interactions between features⁹. These approaches overlook the synergistic effects of combined indices, which imposes limitations on constructing highly reliable predictive models and hinders the development of effective myopia prevention strategies¹⁵. While machine learning (ML) studies¹⁶ have enhanced myopia prediction efficacy¹⁷ by capturing high-order nonlinear relationships between covariates and outcomes, achieving higher accuracy and construct sophisticated models that better adapt to data characteristics¹⁸. Traditional ML still face critical limitations. These models often rely on single algorithms¹⁹ that struggle to accommodate diverse data features, while their lack of interpretability²⁰ further restricts both screening applications and clinical utility.
Therefore, exploring ML-based multi-indicator models can more accurately predict childhood myopia. This study develops an ML model integrating age, AL/CCR, and refractive parameters to identify age-specific thresholds for younger school-aged children, leveraging Shapley Additive explanation (SHAP) for risk assessment. This approach fills a research gap by providing interpretable, age-tailored screening, enabling early identification of at-risk children, timely referral, and potential delay of myopia progression to improve visual outcomes²¹.
This study was conducted from November 2023 to January 2024 and utilized stratified sampling by grade along with probability proportionate to size sampling. The study covered 16 schools and kindergartens across two counties in Changsha City, including senior kindergarten classes and all grades of primary school students. Based on previous myopia survey data, the minimum sample size required for each grade was calculated using the estimation formula for simple random sampling. Notably, children with strabismus, amblyopia, a history of ocular surgery, keratoconus, accommodative spasms, or severe systemic diseases were excluded from the study. Additionally, participants who had contraindications to mydriasis were not included in the study. Ultimately, 2368 children’s data were collected. All participating students joined the survey after providing informed consent. Informed consent was obtained from both participating students and their parents or legal guardians before study commencement. This study has been approved by the Ethics Committee of the Changsha Municipal Center for Disease Control and Prevention and was conducted following the principles of the Declaration of Helsinki.
Uncorrected distance visual acuity (UCDVA) was assessed using a standard logarithmic visual acuity chart at a distance of 5 m, determining an individual’s visual clarity without corrective lenses. Additionally, autorefraction was performed before and after pupil dilation using an automated refractometer (KR-800; Topcon, Japan) to ensure continuity and accuracy of the measurements. All participants were examined by the same optometrist for both pre- and post-dilation autorefraction to maintain consistency in the results.
Ocular biometric parameters were measured by noncontact partial-coherence laser interferometry (IOLMaster 500, Zeiss, Germany), including AL, K1/K2 of CCR, and anterior chamber depth (ACD). Five measurements were taken for each parameter, and the average value was used to ensure data reliability. Before dilation, all children underwent noncontact intraocular pressure measurement and slit-lamp examination to rule out contraindications to dilation. Pupillary dilation was performed using 0.5% tropicamide eye drops administered once every 5 min for a total of four doses, followed by refraction 30 min after the last dose.
Spherical equivalent (SE) was calculated as the sphere power plus half the cylinder power. UCDVA < 5.0 was classified as having poor vision. Screening myopia was defined as UCDVA < 5.0 and non-cycloplegic SE < -0.50 diopters (D). Myopia was diagnosed when cycloplegic SE was ≤ -0.50D. To ensure the quality of the study, all team members received uniform training before the study began, and the equipment was calibrated with standard eyes before formal testing. Moreover, 5% of the ophthalmic examination items were re-tested and verified according to the number of examinees to identify and promptly correct any discrepancies.
The dataset included demographic information (age, gender, height, and weight) for each participant and ocular profile (UCDVA, SE, AL, CCR and ACD) for both eyes. The analysis was limited to measurements from the right eye to maintain data independence and reduce potential bias, ensuring a more accurate and impartial assessment.
Two samples of individuals wearing orthokeratology lenses and one column with outliers were removed from the dependent variable, resulting in a final sample size of 2365 cases.
Missing data were addressed carefully to minimise bias. To address this issue, multiple imputation techniques were employed for the dataset. All variables included had < 20% of their data missing: height (1 case, 0.04%), weight (1 case, 0.04%), UCDVA (10 cases, 0.42%), non-cycloplegic SE (7 cases, 0.3%), cycloplegic SE (81 cases, 3.42%), AL (221 cases, 9.34%), CCR (221 cases, 9.34%), and ACD (221 cases, 9.34%). Assuming the data were missing at random, the fully conditional specification method was utilised, which was available based on five replications and a chained equation approach method. Missing values were imputed using the MICE (Multivariate Imputation by Chained Equations) package (version 3.16.0) for the R programming language.
Correlation analysis was conducted to reduce model bias caused by the interdependence of variables. A heatmap of correlation coefficients was generated to visualize these relationships. To prevent predictive model instability caused by high collinearity among independent variables, variables with inter-independent variable correlation coefficients > 0.7 and relatively lower correlation with the dependent variable were removed.
The diagnostic ability of single indicator variables was visually assessed using ROC curves, and the model’s effectiveness was measured via two key indicators: the area under curve (AUC) and the cutoff point determined by the Youden Index (Sensitivity + Specificity − 1), which was employed to identify the optimal threshold balancing specificity and sensitivity.
In ML, we have adopted Tidymodels (version 1.1.1) for tidy modeling. The model’s predictive accuracy was enhanced by extracting, selecting, and transforming features from the original data through data centering, scaling, zero-variance predictors removal, and dummy variable treatment for categorical variables. Five representative models were selected based on data characteristics: Extreme Gradient Boosting (XGBoost), Regularized Logistic Regression (LR_reg), K-Nearest Neighbor (KNN), Random Forest (RF), and Support Vector Machine (SVM). We divide the dataset into training (70%) and testing (30%) sets to reduce the risk of overfitting and bias. To ensure the robustness of the validation results, we used 5 × 10-fold cross-validation: the dataset was split into 10 subsets, training on 9 and validating on 1 in each fold. Repeating this 5 times with random splits generated 50 results, whose average served as the final performance metric to reduce partitioning bias. Our hyperparameter optimization strategy improved search efficiency through a competitive grid search and increased the possibility of finding the optimal hyperparameter combination. We comprehensively evaluate the performance of multiple variable combinations and predictive models, using techniques such as cross-validation to assess the generalization performance of the models. The key indicator for evaluating model efficacy was the AUC of the ROC curve. Additionally, we used three other indicators: accuracy, brier score, and F1 score, to comprehensively evaluate the performance of the ML algorithms. Using these comprehensive evaluation indicators, we selected the most effective predictive model for further research.
We used R software (version 4.3.3) to conduct all statistical analyses and build multi-indicator predictive models based on ML methods. We used SHAP values to interpret the models, enhancing their interchangeability. Descriptive statistical analyses were performed for both continuous and categorical variables. T-tests were used to compare metric data, and chi-square tests for categorical data. Statistical significance was set at a two-tailed P-value < 0.05. Figure.1 illustrates the workflow of the study.
Flowchart of screening.
The study involved 2,365 children (1184 boys and 1181 girls), aged 5–12 years (mean 7.8 ± 1.7 years). The myopia rate was 11.2% (95% confidence interval [CI]: 9.9%~12.5%), poor vision rate was 34.9% (95%CI: 33.0 ~ 36.8%), and screening myopia rate was 13.4% (95%CI: 12.0 ~ 14.7%). The mean non-cycloplegic SE was − 0.03 ± 1.12 D, mean cycloplegic SE was 0.64 ± 1.23D, mean UCDVA was 4.9 ± 0.2, mean AL was 22.95 ± 0.86 mm, mean CCR was 7.82 ± 0.26 mm, and mean AL/CCR was 2.94 ± 0.09. Significant differences (P < 0.05) were observed in age, grade, height, weight, vision metrics, and ACD between children with myopia and those without. Children with myopia were generally older, taller, and heavier, with longer ALs, higher AL/CCR ratios, and deeper ACDs. Gender distribution showed no significant difference between the two groups (χ² = 0.207, P = 0.207). Compared with children with poor vision, those identified through screening myopia had a higher proportion of myopia (72.5% vs. 28.1%). (Table 1)
We used a pearson correlation matrix to analyse relationships among 10 variables (Fig. 2a). Non-cycloplegic SE strongly correlated with cycloplegic SE (r = 0.82) and AL/CCR (r = 0.75). It exhibited a moderate association with uncorrected distant visual acuity(r = 0.45), and a negligible correlation with CCR(r = 0.04). Furthermore, age exhibited pronounced collinearity with both height and weight, prompting our decision to exclusively incorporate age in subsequent analyses, omitting height and weight.
Given the strong correlation (r > 0.8) between non-cycloplegic SE and cycloplegic SE, we segmented the age variable for a more nuanced subgroup analysis (Fig. 2b), revealing that the effectiveness of SE testing varied significantly with age. The difference in SE before and after cycloplegic refraction decreased with age. This variance was statistically significant in the 5–10 years age group, with a mean difference exceeding 0.50D. In children > 10 years, the difference was not statistically significant (P > 0.05). Figure 2c illustrates the prevalence of myopia among children of different age groups using data obtained through three distinct methods. The prevalence of poor vision follows a U-shaped pattern as children age, decreasing from ages five to seven years and then gradually increasing beyond seven years. Screening myopia mirrors actual myopia rates, with higher screening rates at five years (8.3% vs. 2.5%) and lower rates at 12 years (51.1% vs. 57.4%).
Correlation heat map of visual screening variables (a). Non-cycloplegic spherical equivalent and cycloplegic spherical equivalent diopters (b). The prevalence of myopia according to three methods in different age groups (c). NS non-significant.
The ROC curve (Fig. 3) compared conventional myopia screening methods and ocular biometric measurements. The cutoff values, sensitivity, and specificity of different screening examined. The most effective single-indicator screening diagnostic methods were AL/CCR and screening myopia, with AUCs of 0.919 (95% CI: 0.899 to 0.939) and 0.911 (95% CI: 0.890 to 0.932), respectively. For AL/CCR, the optimal cutoff value was 3.005, with a specificity of 0.826 and sensitivity of 0.895; screening myopia had a specificity of 0.864 and sensitivity of 0.959.
Comparison of receiver operator characteristic curve.
Additionally, we compared UCDVA, poor vision, non-cycloplegic SE, screening myopia, AL, AL/CCR, and ACD with the gold standard across different age groups (Table 2). Non-cycloplegic SE had the highest AUC values among the conventional screening methods across all age groups, while the AUCs of UCDVA and poor vision significantly increased with age. The UCDVA cutoff was below 5.0 for children ≤ 8 years, with younger ages correlating with lower cutoff values. Among the biometric measurements, AL/CCR consistently had higher AUC values than AL, CCR, and ACD, with a stable cutoff value of approximately 3 across all age groups. ACD exhibited lower diagnostic efficacy in all age groups.
We developed four diagnostic models using various ML algorithms, including XGBoost, LR-reg, KNN, RF, and SVM, to accommodate four distinct variable combinations. The selection of the optimal model within the training set is depicted in Fig. 4. Subsequently, we assessed the ultimate performance and generalization capacity of these models on the test set, as detailed in Table 3. The findings revealed that the XGBoost model, using age, UCDVA, non-cycloplegic SE, and AL variables achieved the highest AUC (0.983). The RF model, using age, UCDVA, non-cycloplegic SE, and AL/CCR, showed the highest accuracy (0.975), F1 score (0.868), and the lowest Brier score (0.022).
Optimization and selection of machine learning joint diagnostic models based on multi-indicator combinations.
We used SHAP values to interpret the top-performing XGBoost model (model1) and RF model (model2), illustrating how these variable indicators predict myopia. Figure 5a shows the important features in the XGBoost model, ranked by significance: non-cycloplegic SE, UCDVA, AL, and age. Figure 5b shows the most important features in the RF model, ranked by significance, with key predictors including non-cycloplegic SE, AL/CCR, UCDVA, and age. Figure 5c and d present two case examples (ID = 6), one classified as non-myopia and the other as myopia, to further highlight model’s interpretability.
Model variable importance ranking (top) and Shapley Additive explanation force plot for selected students (bottom). Panels a and c represent the XGBoost model (model1), while Panels b and d represent the RF model (model2).
Our study confirms that age-specific characteristics are crucial for screening and diagnosing myopia. We found that visual acuity and screening myopia are less effective in young children, while non-cycloplegic refraction is more suitable for children > 10 years. Ocular biometric measurements, particularly the AL/CCR, show higher efficacy, with an optimal diagnostic threshold above 3.003. The multi-indicator joint diagnostic model based on ML exhibited stronger predictive accuracy and generalization. The best model for diagnosing myopia using age, UCDVA, non-cycloplegic SE, and AL was XGBoost, with an AUC of 0.983 and an accuracy of 0.970. The RF model using age, UCDVA, non-cycloplegic SE, and the AL/CCR ratio had an AUC of 0.981 and an accuracy of 0.975.
Previous studies have shown that age and years of education are highly correlated with refractive error in children²². Liu et al.¹⁴ noted that non-cycloplegic refractive assessments in children often overestimate myopia owing to accommodation, resulting in frequent misdiagnoses of myopia and hyperopia. Our study also observed differences in myopia rate and SE before and after cycloplegia, particularly in children < 10 years. The differences in SE among age groups are due to stronger accommodative abilities in younger children, which cause shifts in SE during non-cycloplegic refraction²³. The Tehran Eye Study²⁴ reported 99% sensitivity but only 80.4% specificity for non-cycloplegic self-assessment refraction for myopia. The variation in measurements with and without cycloplegia is influenced by age, refractive category, and individual differences.
In this study, myopia rates in children aged 5–6 range from 2.8 to 3.0%, aligning with the consensus that the myopia in children under 6 is typically < 5%^25,26. The efficacy and stability of UCDVA were low and unstable in younger age groups, particularly children ≤ 8 years, with a cutoff value below 5.0. This finds that visual development in children is a gradual process of emmetropisation²⁷during which some children may not achieve the standard far-vision level of 5.0 before the age of 6. Consequently, the level of screened myopia may be overestimated in younger children, making it difficult to achieve China’s target of 3% for myopia prevention and control in this age group by 2030, based on non-cycloplegic diagnostic criteria for screened myopia. Current school-based myopia screening, relying on distant visual acuity measurement and non-cycloplegic autorefraction, has limited value in accurately determining refractive status²⁸.
These findings highlight the need to continuously improve screening tools to ensure accuracy and effectiveness across various developmental stages. Previous studies have explored the use of ocular biometric measurement in predicting myopia^29,30particularly, AL which has demonstrated good predictive value^28,31. Notably, the shift toward myopia and acceleration of AL elongation may be evident up to four years before myopia onset, with similar patterns observed across different ethnic groups². Liu et al.¹⁴ suggested that the AL/CCR has a stronger correlation with myopia than AL or CR alone. In our study, the AL/CCR excelled in single-indicator screening, surpassing myopia screening, UCDVA, AL, and ACD. The ROC curve cutoff values were consistently around 3 across age groups, indicating that the AL/CCR can serve an alternative indicator for identifying preschool children with low hyperopia reserves and myopia, aiding in early detection.
Traditional predictive models, including linear regression¹²logistic regression³²cox proportional hazards regression³³and generalized estimating equations. ³⁴, typically rely on statistically significant variables, with baseline SE dominating myopia prediction. For instance, zadnik et al.³⁵ reported that baseline SE achieved an AUC of 0.88, and even combined with axial length/corneal curvature, AUC only marginally improved to 0.893, reflecting overdependence on single parameters. In contrast, ML excels in high-dimensional data processing and longitudinal prediction. A recent study³¹ comparing five ML algorithms (RF, SVM, GBDT, CatBoost, logistic regression) in children aged 6–13 found CatBoost achieving AUC = 0.951 (vs. logistic regression AUC = 0.739), demonstrating ML’s superiority in leveraging complex feature interactions.
Multi-model approaches outperform single models¹⁹due to data complexity, algorithm heterogeneity, and diverse clinical needs. For example, in pediatric myopia datasets³⁶orthogonal matching pursuit (OMP) excelled in SE prediction, while kernel ridge (KR) and multilayer perceptron (MLP) dominated AL estimation. In this study, five representative algorithms—logistic regression, XGBoost, KNN, RF, and SVM—were selected to capture distinct data patterns and age-specific trends. This complementary integration enhanced adaptability across clinical scenarios. Results identified XGBoost and RF as top performers in multi-indicator models, validating algorithmic diversity as a robustness enhancer.
Ocular biometric measurement methods, while superior⁹ to traditional detection methods, have limitations in monitoring myopia progression due to inconsistent relationships between the AL/CCR and myopia severity. Multi-indicator joint diagnosis is crucial in addressing these limitations, and ML¹⁷ offers promising solutions, particularly through cross-validation and hyperparameter tuning, which can fully utilize data, reduce the risk of overfitting, and improve model adaptability^37,38. Although non-cycloplegic measurements may not reliably determine individual refractive errors, proper data modelling can help classify and identify risk groups for cycloplegic refraction²⁸. Liu et al.³⁹ recommended using the AL/CCR or combining AL with non-cycloplegic auto-refraction for higher accuracy in preschool children. A three-year retrospective study covering 13 cities found that⁴⁰ age, uncorrected distant visual acuity, and SE were predictive factors for high myopia in school-age children, with the RF algorithm achieving an accuracy rate of 0.948 and an AUC of 0.975. Du et al.⁴¹ reported that the AdaBoost model predict refractive status more accurately than direct non-cycloplegic SE estimation, with an 81.7% accuracy rate and 75.2% of SE prediction errors < 0.50D. The reduced effectiveness of screening myopia compared to UCDVA and SE before cycloplegia may be due to the loss or simplification of some information caused by converting continuous variables into categorical variables, affecting the precision of data analysis and the effectiveness of the test.
This study used SHAP values⁴² to demystify the decision-making processes of ML models, particularly XGBoost and RF. SHAP values for each feature variable in the test dataset revealed their contributions to the prediction outcomes. The overall feature importance plot provided an average assessment of each feature’s contribution to the overall predictive results. For personalized risk prediction, the SHAP force plot demonstrated how various features influence individual risk predictions¹⁸. Although individual risk predictions aligned with overall feature importance, variations in specific indicator highlight the heterogeneity among individuals.
Our study had some limitations. First, we used 0.5% tropicamide as the cycloplegic agent rather than cyclopentolate, considered the gold standard. Second, the endpoint of cycloplegia was based on examiner records without objective measurement standards. Therefore, complete cycloplegia cannot be confirmed in all cases. Third, the study was not cohort-based and lacked a temporal correlation between refractive error and the AL/CCR, leaving room for observational and inclusion bias. Fourth, the predictive model did not include influencing factors⁴³.Baseline ocular biometrics and refractive error, while predictive of incident myopia, also reflect previous risky behaviours, which may have led to underestimating the potential benefits of behavioral changes. This may lead to an underestimation of the potential benefits that can be obtained from behavioural changes. Finally, this study did not undergo rigorous external validation. Future studies should incorporate multicenter validation to enhance the reliability and generalizability of the model.
Our study highlights the limitations of traditional vision tests in screening for myopia in children, particularly their unsatisfactory detection efficiency. Age specificity is crucial when diagnosing and screening for myopia at different developmental stages. The AL/CCR demonstrated superior performance and can serve as a single indicator for identifying myopia in children, while non-cycloplegic SE is more suitable for older children. Our ML-based multi-indicator joint diagnostic model enhances diagnostic accuracy and practical applicability. The interpretability of SHAP values allows for more in-depth group predictions and individual myopia diagnoses.
The datasets generated and analyzed during the current study are not publicly available because of involving students’ privacy but are available from the corresponding author upon reasonable request.
Sankaridurg, P. et al. IMI impact of myopia. Investig. Ophthalmol. Vis. Sci. 62 https://doi.org/10.1167/iovs.62.5.2 (2021).
Wolffsohn, J. S. et al. IMI – Myopia control reports overview and introduction. Investig. Ophthalmol. Vis. Sci. 60, M1–m19. https://doi.org/10.1167/iovs.18-25980 (2019).
Article  Google Scholar 
Holden, B. A. et al. Global prevalence of myopia and high myopia and Temporal trends from 2000 through 2050. Ophthalmology 123, 1036–1042. https://doi.org/10.1016/j.ophtha.2016.01.006 (2016).
Article  PubMed  Google Scholar 
Kandel, H., Khadka, J., Goggin, M. & Pesudovs, K. Impact of refractive error on quality of life: a qualitative study. Clin. Exp. Ophthalmol. 45, 677–688. https://doi.org/10.1111/ceo.12954 (2017).
Article  PubMed  Google Scholar 
Huang, J., Ma, W., Li, R., Zhao, N. & Zhou, T. Myopia prediction for children and adolescents via time-aware deep learning. Sci. Rep. 13, 5430. https://doi.org/10.1038/s41598-023-32367-0 (2023).
Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 
Ma, Y. et al. Cohort study with 4-year follow-up of myopia and refractive parameters in primary schoolchildren in Baoshan district, Shanghai. Clin. Exp. Ophthalmol. 46, 861–872. https://doi.org/10.1111/ceo.13195 (2018).
Article  PubMed  PubMed Central  Google Scholar 
Morgan, I. G. et al. IMI risk factors for myopia. Investig. Ophthalmol. Vis. Sci. 62 https://doi.org/10.1167/iovs.62.5.3 (2021).
Colpa, L. et al. Nonsurgical consecutive Exotropia following childhood esotropia: A multicentered study. Am. J. Ophthalmol. 258, 130–138. https://doi.org/10.1016/j.ajo.2023.07.021 (2024).
Article  PubMed  Google Scholar 
Jong, M., Sankaridurg, P., Naduvilath, T. J., Li, W. & He, M. The relationship between progression in axial length/corneal radius of curvature ratio and spherical equivalent refractive error in myopia. Optometry Vis. Science: Official Publication Am. Acad. Optometry. 95, 921–929. https://doi.org/10.1097/opx.0000000000001281 (2018).
Article  Google Scholar 
Liu, S. et al. Cutoff values of axial length/corneal radius ratio for determining myopia vary with age among 3–18 years old children and adolescents. Graefe’s Archive Clin. Experimental Ophthalmol. = Albrecht Von Graefes Archiv fur Klinische Und Experimentelle Ophthalmologie. 262, 651–661. https://doi.org/10.1007/s00417-023-06176-0 (2024).
Article  Google Scholar 
Jonas, J. B. et al. IMI prevention of myopia and its progression. Investig. Ophthalmol. Vis. Sci. 62 https://doi.org/10.1167/iovs.62.5.6 (2021).
Chua, S. Y. et al. Age of onset of myopia predicts risk of high myopia in later childhood in myopic Singapore children. Ophthalmic Physiological Optics: J. Br. Coll. Ophthalmic Opticians (Optometrists). 36, 388–394. https://doi.org/10.1111/opo.12305 (2016).
Article  Google Scholar 
Ghorbani Mojarrad, N., Williams, C. & Guggenheim, J. A. A genetic risk score and number of myopic parents independently predict myopia. Ophthalmic Physiological Optics: J. Br. Coll. Ophthalmic Opticians (Optometrists). 38, 492–502. https://doi.org/10.1111/opo.12579 (2018).
Article  Google Scholar 
Liu, L. et al. Prediction of premyopia and myopia in Chinese preschool children: a longitudinal cohort. BMC Ophthalmol. 21, 283. https://doi.org/10.1186/s12886-021-02045-8 (2021).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Guo, X. et al. Noncycloplegic compared with cycloplegic refraction in a Chicago School-Aged population. Ophthalmology 129, 813–820. https://doi.org/10.1016/j.ophtha.2022.02.027 (2022).
Article  PubMed  Google Scholar 
Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. Jama 319, 1317–1318. https://doi.org/10.1001/jama.2017.18391 (2018).
Article  PubMed  Google Scholar 
Zhang, J. & Zou, H. Artificial intelligence technology for myopia challenges: A review. Front. Cell. Dev. Biology. 11, 1124005. https://doi.org/10.3389/fcell.2023.1124005 (2023).
Article  Google Scholar 
Tong, H. J. et al. Machine learning to analyze the factors influencing myopia in students of different school periods. Front. Public. Health. 11, 1169128. https://doi.org/10.3389/fpubh.2023.1169128 (2023).
Article  PubMed  PubMed Central  Google Scholar 
Yang, X. et al. Prediction of myopia in adolescents through machine learning methods. Int. J. Environ. Res. Public. Health. 17 https://doi.org/10.3390/ijerph17020463 (2020).
Lin, H. et al. Prediction of myopia development among Chinese school-aged children using refraction data from electronic medical records: A retrospective, multicentre machine learning study. PLoS Med. 15, e1002674. https://doi.org/10.1371/journal.pmed.1002674 (2018).
Article  PubMed  PubMed Central  Google Scholar 
Han, X., Liu, C., Chen, Y. & He, M. Myopia prediction: a systematic review. Eye (London England). 36, 921–929. https://doi.org/10.1038/s41433-021-01805-6 (2022).
Article  PubMed  Google Scholar 
Morgan, I. G. & Jan, C. L. China turns to school reform to control the myopia epidemic: A narrative review. Asia-Pacific J. Ophthalmol. (Philadelphia Pa). 11, 27–35. https://doi.org/10.1097/apo.0000000000000489 (2022).
Article  Google Scholar 
He, X. et al. Normative data and percentile curves for axial length and axial length/corneal curvature in Chinese children and adolescents aged 4–18 years. Br. J. Ophthalmol. 107, 167–175. https://doi.org/10.1136/bjophthalmol-2021-319431 (2023).
Article  PubMed  Google Scholar 
Fotouhi, A., Morgan, I. G., Iribarren, R., Khabazkhoob, M. & Hashemi, H. Validity of noncycloplegic refraction in the assessment of refractive errors: the Tehran eye study. Acta Ophthalmol. 90, 380–386. https://doi.org/10.1111/j.1755-3768.2010.01983.x (2012).
Article  PubMed  Google Scholar 
Dirani, M. et al. Prevalence and causes of decreased visual acuity in Singaporean Chinese preschoolers. Br. J. Ophthalmol. 94, 1561–1565. https://doi.org/10.1136/bjo.2009.173104 (2010).
Article  CAS  PubMed  Google Scholar 
Lan, W. et al. Refractive errors in 3–6 year-old Chinese children: a very low prevalence of myopia? PLoS One. 8, e78003. https://doi.org/10.1371/journal.pone.0078003 (2013).
Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 
Schein, Y., Yu, Y., Ying, G. S. & Binenbaum, G. Emmetropization during early childhood. Ophthalmology 129, 461–463. https://doi.org/10.1016/j.ophtha.2021.11.021 (2022).
Article  PubMed  Google Scholar 
Sankaridurg, P. et al. Comparison of noncycloplegic and cycloplegic autorefraction in categorizing refractive error data in children. Acta Ophthalmol. 95, e633–e640. https://doi.org/10.1111/aos.13569 (2017).
Article  CAS  PubMed  PubMed Central  Google Scholar 
Sanz Diez, P., Yang, L. H., Lu, M. X., Wahl, S. & Ohlendorf, A. Growth curves of myopia-related parameters to clinically monitor the refractive development in Chinese schoolchildren. Graefe’s Archive Clin. Experimental Ophthalmol. = Albrecht Von Graefes Archiv fur Klinische Und Experimentelle Ophthalmologie. 257, 1045–1053. https://doi.org/10.1007/s00417-019-04290-6 (2019).
Article  Google Scholar 
Xiao, J. et al. Analysis and modeling of myopia-related factors based on questionnaire survey. Comput. Biol. Med. 150, 106162. https://doi.org/10.1016/j.compbiomed.2022.106162 (2022).
Article  PubMed  Google Scholar 
Peng, W. et al. Does multidimensional daily information predict the onset of myopia? A 1-year prospective cohort study. Biomed. Eng. Online. 22 https://doi.org/10.1186/s12938-023-01109-8 (2023).
French, A. N., Morgan, I. G., Mitchell, P. & Rose, K. A. Risk factors for incident myopia in Australian schoolchildren: the Sydney adolescent vascular and eye study. Ophthalmology 120, 2100–2108. https://doi.org/10.1016/j.ophtha.2013.02.035 (2013).
Article  PubMed  Google Scholar 
Wang, S. K. et al. Incidence of and factors associated with myopia and high myopia in Chinese children, based on refraction without cycloplegia. JAMA Ophthalmol. 136, 1017–1024. https://doi.org/10.1001/jamaophthalmol.2018.2658 (2018).
Article  PubMed  PubMed Central  Google Scholar 
Zhang, M. et al. Validating the accuracy of a model to predict the onset of myopia in children. Investig. Ophthalmol. Vis. Sci. 52, 5836–5841. https://doi.org/10.1167/iovs.10-5592 (2011).
Article  Google Scholar 
Zadnik, K. et al. Ocular predictors of the onset of juvenile myopia. Investig. Ophthalmol. Vis. Sci. 40, 1936–1943 (1999).
CAS  Google Scholar 
Zhu, S. et al. Prediction of spherical equivalent refraction and axial length in children based on machine learning. Indian J. Ophthalmol. 71, 2115–2131. https://doi.org/10.4103/ijo.Ijo_2989_22 (2023).
Article  PubMed  PubMed Central  Google Scholar 
Wang, Y. et al. Machine learning models for predicting Long-Term visual acuity in highly myopic eyes. JAMA Ophthalmol. 141, 1117–1124. https://doi.org/10.1001/jamaophthalmol.2023.4786 (2023).
Article  PubMed  PubMed Central  Google Scholar 
Li, W. et al. Study of myopia progression and risk factors in Hubei children aged 7–10 years using machine learning: a longitudinal cohort. BMC Ophthalmol. 24, 93. https://doi.org/10.1186/s12886-024-03331-x (2024).
Article  PubMed  PubMed Central  Google Scholar 
Li, S. M. et al. Annual incidences and progressions of myopia and high myopia in Chinese schoolchildren based on a 5-Year cohort study. Investig. Ophthalmol. Vis. Sci. 63, 8. https://doi.org/10.1167/iovs.63.1.8 (2022).
Article  Google Scholar 
Guan, J. et al. Prevalence patterns and onset prediction of high myopia for children and adolescents in Southern China via Real-World screening data: retrospective School-Based study. J. Med. Internet. Res. 25, e39507. https://doi.org/10.2196/39507 (2023).
Article  PubMed  PubMed Central  Google Scholar 
Du, B. et al. Prediction of spherical equivalent difference before and after cycloplegia in school-age children with machine learning algorithms. Front. Public. Health. 11, 1096330. https://doi.org/10.3389/fpubh.2023.1096330 (2023).
Article  PubMed  PubMed Central  Google Scholar 
Zhao, J. et al. Development and validation of predictive models for myopia onset and progression using extensive 15-year refractive data in children and adolescents. J. Translational Med. 22, 289. https://doi.org/10.1186/s12967-024-05075-0 (2024).
Article  Google Scholar 
Tideman, J. W. L., Polling, J. R., Jaddoe, V. W. V., Vingerling, J. R. & Klaver, C. C. W. Environmental risk factors can reduce axial length elongation and myopia incidence in 6- to 9-Year-Old children. Ophthalmology 126, 127–136. https://doi.org/10.1016/j.ophtha.2018.06.029 (2019).
Article  PubMed  Google Scholar 
Download references
The authors thank Han Xun and Haixiang Zhou for their assistance with this study.
Supported by Health Research Project of Hunan Provincial Health Commission (grant number: W20243235).
Changsha Municipal Center for Disease Control and Prevention, No. 509, Wanjiali Second North Road, Kaifu District, Changsha, 410001, Hunan, China
Qi Feng, Xin Wu, Qianwen Liu, Yuanyuan Xiao, Xixing Zhang & Yan Chen
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Q.F. designed the study, contributed the statistical analysis, interpreted the results and wrote the original draft. Q.W. L., X.W. and Y.Y.X contributed data curation, software, and materials/ analytic. X.X.Z. and Y.C. contributed designed, administration, and supervision. All authors reviewed and approved the manuscript prior to submission.
Correspondence to Yan Chen.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Feng, Q., Wu, X., Liu, Q. et al. Interpretable machine learning models for predicting childhood myopia from school-based screening data. Sci Rep 15, 19811 (2025). https://doi.org/10.1038/s41598-025-05021-0
Download citation
Received: 03 December 2024
Accepted: 30 May 2025
Published: 05 June 2025
DOI: https://doi.org/10.1038/s41598-025-05021-0
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2025 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

source