AI enhanced optimization of college physical education programs using hybrid genetic algorithms and learning based fitness evaluation.

AI enhanced optimization of college physical education programs using hybrid genetic algorithms and learning based fitness evaluation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 11;16:16967. doi: 10.1038/s41598-026-45357-9 AI enhanced optimization of college physical education programs using hybrid genetic algorithms and learning based fitness evaluation Xinlong Wang 1 , Wenze Sun 2, ✉ Author information Article notes Copyright and License information 1 Department of Physical Education, Beijing Wuzi University, Beijing, 101149 China 2 Department of Physical Education, Ocean University of China, Qingdao, 266100 China ✉ Corresponding author. Received 2025 Jul 20; Accepted 2026 Mar 18; Collection date 2026. © The Author(s) 2026 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/ . PMC Copyright notice PMCID: PMC13230744  PMID: 41965788 Abstract Structured physical education (PE) programs are essential for fostering students’ physical development, cognitive performance, and emotional well-being in academic settings. This study introduces a novel intelligent decision-making algorithm (IDA) for the dynamic optimization and real-time assessment of college PE programs. The proposed framework synergistically integrates a hybrid genetic algorithm (GA) that is viable global optimizer with pattern search (PS) that is an adaptive local search technique. The exploration in the search domain is performed with GA while PS exploitation in a minimal computational budget is performed by the PS. The system encodes program performance data into chromosome-like representations, enabling nuanced evaluation across academic progress, health indices, and affective-psychomotor domains. A tailored vector of weight factors refines the fitness function to reflect individual learning trajectories and institutional goals. Experimental results demonstrate that the model achieves a high prediction accuracy of 98%, with quantifiable improvements of 151.13% in holistic performance, 19.63% in educational metrics, and 26.7% in psychomotor development as compared with reported results. Comparative models, including Random Forest (RF), Adaptive Neuro-Fuzzy Inference System (ANFIS), and RF regression, that achieved accuracies of 88.21%, 94.49%, and 82%, respectively. The hybrid framework maintained a mean global fitness value of 5.3451 × 10⁻¹² with an average computational time of 1357.04 s over 100 runs that is used to validate the reliability of the proposed framework. By enhancing efficiency and reducing computational complexity, this AI-driven evaluation model offers a scalable and intelligent approach for real-time optimization and policy refinement in higher education PE curriculum planning. Keywords: College physical education programs, Evolutionary computation, Pattern search technique, Intelligent hybrid optimization, Monte Carlo simulations Subject terms: Computational biology and bioinformatics, Engineering, Mathematics and computing Introduction In the previous decade, computing technology has changed college students’ academic performance, morals, and social obligations. Effective physical education training (PET) for college students is necessary to promote academic achievement, health, and social and moral responsibility. Recent trends in physical education indicate a shift from a focus solely on physical fitness to a more holistic approach that emphasizes comprehensive student development. More schools want to upgrade their PE programs. Institutions must find better ways to evaluate training and student performance. Intelligent Decision Algorithms (IDAs) will improve decision-making using data and weighted parameters. This study implements an intelligent algorithm to analyze and optimize the college PET, which affects student learning, performance, and well-being on standard parameters 1 . In recent years, data-driven education research has grown in prominence. Most qualitative and subjective assessments in the classic evaluation paradigm produce contradictory and skewed results on an enhanced program. IDAs can be added to the evaluation framework to help educators and administrators quantify program effectiveness 2 . This strategy improves evaluation methods and helps show how different situations effect pupils in physical education. Physical education aims include curriculum disruption, instructor quality, and student engagement 3 . With diverse approaches, IDAs can identify correlations and trends in big data sets that thematic assessment cannot. For instance, comparing student demographics to their achievements will reveal trends of physical education class engagement. It can also help ensure that educational programs meet student needs, improving school environments 4 . The educational foundation of this study aligns with Bloom’s Taxonomy, which outlines three essential domains of learning: cognitive (knowledge and intellectual skills), affective (attitudes and emotions), and psychomotor (physical skills). Our performance indicators are educational performance (EP), affective domain (AD), and psychomotor domain (PD) directly correspond to these dimensions. Moreover, the adaptive nature of the proposed IDA framework is informed by Constructivist Learning Theory, which advocates personalized and experiential learning based on individual performance data. This theoretical grounding supports the integration of intelligent decision-making algorithms into educational design and evaluation. IDAs go beyond assessment since they can improve physical education programs. In this approach, institutions can collect data year-after-year and evaluate curricular innovations and teaching methods on pupils. This method helps identify best practices that should be repeated and areas that require improvement to ensure physical education programs suit students’ needs 5 . Integration of IDAs can assist meet calls for cardinal and evidence-based approaches to college student engagement and aerobic activity. In conclusion, using IDAs to evaluate college physical education programs will improve student performance and training efficacy. Mass data analysis helps educators evaluate student involvement and performance to make better judgments. This project will hybridize two artificial intelligence-based methodologies to create an intelligent decision algorithm (IDA). Genetic algorithm (GA) is a global optimizer, and pattern search (PS) uses intelligent local search to optimize student dataset fitness function weight vectors. The projected IDA’s key contributions are: The IDA framework was developed by combining GA as a global optimizer and PS as an intelligent local search strategy to transform decision parameters into an equal-length decoded chromosome. The proposed IDA is compared to state-of-the-art methods such as SVM, NB, GRNN 6 , RF hybridized with ANFIS 7 , and RF Regression 8 on performance indicators such as overall, educational, health, cognitive, affective, and psychomotor domains, and computational complexity in terms of time. To improve the evaluation of physical education trainings on student performance and health, a private dataset of 260 college students was added to the Kaggle dataset. A huge number of independent Monte Carlo simulations ensures the application of the suggested IDA, dependability, and computational complexity in terms of time. Section II provide a comprehensive literature survey along while the material and method section, the dataset, characteristics, IDA framework, GA, and PS are detailed. Section IV presents the suggested approach’s parameter values and settings, hardware requirements, results, discussion, and comparison to state-of-the-art techniques. The conclusion and future work are discussed in the final part. Related work and state of art Recent developments in educational and AI-based optimization research have revealed the prophetic prospect of hybrid models and intelligent systems while improving different perspectives in academic and personality development. A case in point is a study that demonstrated the role of perceived organizational support in predicting career adaptability among university students, suggesting that adjustment of educational environments should be made for better results 9 . Another set of state-of-the-art deep learning models, TASTA and ResLNet have also been presented by combining both spatial and temporal attention mechanisms for better video Q&A and AR performance, demonstrating great flexibility of AI in distrusting sophisticated educational tasks 10 , 11 . Such hybrid AI models are now more often exploited to build adaptive learning environment, such as those in automatic detection of social support in online forums for reduced burnout during learning by enhanced self-regulation 12 . Furthermore, decision-making process optimization models highlight the use of AI in increasing efficiency and accuracy in a number of industries such as education, product design etc 13 , 14 . A number of studies have examined the beneficial effects of PE in academic and social development of students. For example, it has been reported that exercise is directly related to academic achievement and health and the importance of adding a physical education as a daily aspect of academics 15 . Vast research also reveal that daily physical activity is essential for superior cognitive functions and mental health that inevitably leads to increased academic focus and achievement 16 . Also, team play and cooperation games encourage socialemotional development with this essential for the student maturity 17 . These results are consistent with previous investigations on hybrid deep learning models in order to enhance skin lesion segmentation, confirming the promising role of AI for fine-tuning personalized educational and health-intervention approaches 18 . In addition, research has also added to the growing understanding by investigating AI-driven models in health and precision prediction that can be adopted for real-time optimization in education 19 , 20 . Furthermore, predictive models for disease detection and skin lesion analysis represent a model for intelligent framework development that can adapt or actively optimize PE programs as the individual leans and info on health improves 21 , 22 . Through the application of AI models to education, it is possible for institutions to establish personalized learning environments that enhance academic outcomes, as well as safeguarding students’ mental and physical health. Many scholars have examined the effects of physical education on students’ activities, class behavior, test scores, communication skills, and ethics 23 . Trudeau et al. 24 found a link between school and college physical activities and student health and academic success. The study also supports quasi-experimental data that PET should be done a few hours a day. Donnelly & Lambourne 25 examined the effects of physical activity on student achievement and found that students who engaged in this type of physical activity had better records and better mental health, including better focus and memory. These results show that PET is not just an extracurricular activity but a key part of child education that improves cognitive and academic skills through active engagement. McKenzie et al. 26 stated that team sports and cooperative activities improve peer belonging, interpersonal relations, social-emotional development, and life skills. It also highlights how cooperative learning activities foster social-emotional growth, improved mental and physical health. This suggests that physical education promotes a healthy lifestyle, which is essential for student development. However, the primary issue in the aforementioned references lies in the intelligent balancing of these activities, as no formal frameworks or architectures are evident for a specific set of datasets. This limitation is attributed to traditional data analysis tools, which do not accommodate nonlinear constraints optimization for large volumes of data and their adaptive variables. The complexity of the procedures and the multitude of variables involved render AI a feasible solution for accelerating and improving the precision of PET evaluations. Modern machine learning (ML) techniques offer promising avenues to enhance student outcomes by personalizing physical education programs to better align with individual learning needs. Initial studies aimed to employ machine learning algorithms to conduct a detailed analysis of student-related data, such as the frequency of physical activities and other relevant metrics, with the objective of predicting student performance 27 . The researchers employed regression models and classification algorithms to assess the impact of factors such as attendance and participation on student fitness gains. The initial applications demonstrated that prioritizing data analysis could enhance instructional processes, resulting in improved physical education curricula tailored to each student’s potential 28 . The application of machine learning techniques in monitoring and evaluating student performance facilitates the improvement of training models used in physical education. Research indicates that machine learning can analyze data from wearable fitness trackers to monitor students’ performance metrics in real-time, including heart rate, activity level, and skill competence, thereby providing immediate feedback to both students and instructors 29 . This real-time data analysis engaged students by showcasing improvements and provided trainers the capability to adjust training schedules accordingly. Initial applications of machine learning in physical education established a basis for enhancing training frameworks that are empirically supported to improve student performance and engagement in physical education activities 30 – 32 . The advancement of multiple machine learning techniques, including support vector machines (SVM), k-nearest neighbors (KNN), naive Bayes (NB) algorithm, Q-learning epileptic seizure classifier, JAYA-optimized fuzzy reinforcement learning classifier, probabilistic neural networks (PNN), fuzzy lattices and Neural Reinforcement Learning, and random forests (RF), has demonstrated their relevance across diverse fields such as business, economics, and social sciences 33 – 36 . In 2021, Qingxiang Xu et al. 6 utilized an enhanced random forest (RF) method to enhance student performance by optimizing PET hours and empirical teaching through the dimensional reduction of students’ multidisciplinary achievements in physical education. The RF technique achieved an accuracy of 88.55% and a precision of 88.21%. While the study in 20 addresses several critical aspects of physical education quality, it fails to consider significant parameters such as daily mobile phone usage, age, health ratings, student attendance records, and participation in various events. In 2023, Mojsilović et al. developed a combination of the RF algorithm and adaptive neuro-fuzzy inference system (ANFIS) to optimize key mutually independent factors critical to student success, providing guidelines for teachers to enhance their teaching outcomes 7 . The proposed model demonstrates a reliability of 94.49%. However, a significant limitation is that the implementation of RF-ANFIS necessitates a pre-examination of the data and input-output patterns, which may introduce bias. In a recent study 8 , Doz et al. propose a framework utilizing hybrid learning, which integrates fuzzy logic with the RF regression method to examine the impact of demographic factors on college students’ grades. The study emphasizes that the relative impact of various student parameters, such as gender and origin, is less significant on student performance compared to factors like school type, type of physical education training, time spent in physical education, self-study hours, age, and average daily mobile usage duration. Table 1 presents the state of the art against which the results of the current work are compared, detailing the methodologies proposed by various researchers and the corresponding levels of accuracy achieved by their approaches. Table 1. State of art work for comparison with the proposed IDA. Related work Methodology Accuracy (%) Short coming Qingxiang Xu et al. (2021) 6 Random Forest 88.21 The accuracy and precision of the results are compromised. The computational cost of proposed technique is not provided. Mojsilović et al. (2023) 7 RF hybrid with ANFIS 94.49 The training time is quite high, moreover the key parameters are missing. Doz et al. (2023) 8 Random Forest Regression 82 The accuracy is not promising. Qingxiang Xu et al. (2021) 6 SVM, NB and GRNN Method 84.64,84.36 and 84.35 The accuracy and precision is lesser. Open in a new tab This article discusses the transformation of weighted values of various qualitative factors related to PET for school and college students into chromosomes. These chromosomes are subsequently fused using the proposed IDA technique to enhance PET performance in terms of educational score, health indicator, cognitive domain, affective domain, and psychomotor domain. The dataset utilized in this study comprises an additive combination of data from Kaggle 37 and a private dataset involving 260 college-level students. The results of the proposed scheme are compared with the RF algorithm, SVM, NB techniques, and other state-of-the-art methods reported in the literature survey across various performance indicators. The proposed framework’s reliability, stability, and computational complexity in terms of time are validated through 100 independent runs. The citations referenced in this section were collected and annotated using a structured literature review methodology. Searches were conducted across Scopus, IEEE Xplore, and Google Scholar using combinations of keywords such as “physical education,” “machine learning,” “genetic algorithms,” “ANFIS,” and “student performance evaluation.” The identified literature was filtered based on relevance, recency, and technical depth. Each selected article was annotated and thematically categorized into three domains: traditional evaluation techniques, ML-based frameworks for educational improvement, and hybrid AI optimization models. This annotation helped distill core trends, identify performance gaps in existing methods, and justify the novelty and relevance of the proposed IDA model. Unlike prior hybrid models such as RF-ANFIS which combine rule-based fuzzy inference with ensemble learning to enhance predictive accuracy, the proposed hybrid GA-PS introduces a fundamentally different optimization philosophy. Instead of relying on pre-defined fuzzy rules or tree-based structures, our method performs direct continuous optimization of weight vectors using a two-stage search mechanism. GA provides extensive global exploration across the high-dimensional weight space, while PS performs deterministic, derivative-free local refinement to escape shallow local minima and improve convergence stability. This combination creates an adaptive and fully data-driven optimization engine that does not require handcrafted rule sets, prior assumption of input-output relations, or pre-structured model components. To the best of our knowledge, such a GA-initialized pattern search hybrid has not been previously applied in the context of multi-domain educational or physical performance assessment. In addition to ML-based frameworks, some educational decision systems rely on rule-based or lexicon-based evaluation mechanisms, where student responses are scored using predefined rubrics or weight tables developed by experts. While such methods offer interpretability, they are often static and lack adaptability to student diversity and contextual variation. Moreover, these systems are less effective in capturing interactions among multi-dimensional quantitative data. As such, they are not directly suitable for optimizing performance-driven fitness functions like the one proposed in our IDA framework. Several studies have explored the role of pedagogical theories in shaping physical education outcomes. Materials and methods This section comprises three stages: the first stage presents the details of the dataset utilized in the study, the second stage formulates the governing equation for the fitness function, and the third stage outlines the proposed IDA framework, which includes the hybridization of GA and PS, as well as the performance metrics and logical steps undertaken during the optimization of the weight factor vector. Figure 1 illustrates the graphical workflow of the proposed methodology while schematic flow for data preprocessing is provided in Fig. 2 . Fig. 1. Open in a new tab Graphical workflow of proposed IDA framework. Fig. 2. Open in a new tab Schematic flow for data preprocessing for proposed framework. The dataset is transformed into an individual, referred to as a chromosome, with a length corresponding to the number of parameters in the dataset. The decoded chromosome undergoes processing through the designed IDA scheme, which integrates the hybridization of genetic algorithms and an effective local search technique from pattern search. The widely distributed population is initialized by taking into account the lower and upper bounds in the dataset, which facilitates randomization. Consequently, the IDA scheme will be sufficiently capable of identifying the optimal weight from all areas of the search space. The results of the genetic algorithm are integrated with pattern search for future tuning. Once the optimal chromosome is identified, the procedure is repeated for 100 independent runs to conduct basic statistical analyses, ensuring the reliability and stability of the proposed approach. Dataset details The public dataset comprises 81 male and 19 female students sourced from Kaggle 37 . Additionally, to evaluate the efficacy and insights of the PET program, a private dataset from a college containing 260 students is utilized, consisting of 60 females and 200 males. The Kaggle dataset contains 18 attributes, while the private dataset comprises 12 attributes. A total of 20 attributes have been analyzed to assess the impact of PETs on individual education and health outcomes. The primary characteristics of these 20 attributes include gender, daily mobile phone usage, health risks, beneficial subjects, cardiovascular scale, health rating, attendance record, number of events participated in by individual students, weekly student activities, academic profiles encompassing EP, CD, AD, PD, and skill proficiency scores, instructor effectiveness, and student satisfaction, among others. All 20 features from the combined dataset were analyzed since they directly affect physical education evaluation criteria like academic achievement, physical health, cognitive engagement, and program satisfaction. The datasets were preprocessed to remove missing entries, standardize categorical labels, and normalize numerical values using min-max scaling to assure fitness function homogeneity before optimization. We avoided explicit dimensionality reduction methods like PCA and LDA to retain feature interpretability and contribution. Instead, the suggested IDA framework dynamically adjusts attribute weights during optimization. Lower weights for less relevant features reduce their impact on the outcome and serve as an implicit feature selection mechanism in the model’s learning process. To promote reproducibility while adhering to data privacy constraints, an anonymized version of the private dataset containing representative samples has been prepared. This sample retains the original data structure, feature definitions, and statistical characteristics necessary for model evaluation and is available upon request. For interpretability, we categorized the outcome of student performance improvements into three categories: positive (greater than 10% improvement), neutral (0–10% improvement), and negative (decline in score). Based on this classification across the full dataset (including both academic and health indicators), approximately 78% of students showed positive improvement, 16% fell into the neutral range, and 6% experienced a decline in at least one evaluated domain. These categories were not used for training but were derived post hoc to provide a qualitative interpretation of the model’s effect. While most of the dataset variables were derived from structured records and quantitative assessments, certain subjective indicators such as instructor effectiveness and student satisfaction were obtained through standardized feedback surveys. To ensure labeling quality, a dual-annotation process was followed in which two independent annotators evaluated each subjective entry. Discrepancies were resolved through discussion, guided by a third expert reviewer. Inter-annotator agreement was assessed using Cohen’s Kappa statistic, with agreement scores exceeding 0.82 across all subjective fields, indicating high annotation reliability. Most dataset variables were directly obtained from academic and health records, eliminating the need for manual annotation. However, subjective attributes such as ‘instructor effectiveness’ and ‘student satisfaction’ were derived from feedback surveys. These responses were annotated independently by two reviewers using a predefined rubric, and discrepancies were resolved by domain experts. Cohen’s Kappa statistic was used to assess inter-annotator agreement, yielding scores above 0.82, reflecting substantial annotation consistency. Before applying the IDA framework, all numerical features in the dataset were normalized using min-max normalization to scale values into the [0, 1] range. This normalization was performed independently for each training fold to prevent data leakage. Categorical variables like gender, beneficial subjects were first label-encoded into integer values before normalization. This preprocessing step ensured that features with different ranges and magnitudes contributed equitably during optimization and model training. Fitness function creation and performance matrics The performance indicators incorporated into the fitness function OP, EP, HI, CD, AD, and PD were selected based on pedagogical theories such as Bloom’s Taxonomy, which emphasizes multi-dimensional learning outcomes. These indicators reflect cognitive, affective, and psychomotor development and are supported by previous research on the impact of physical education on academic and health performance. Initial correlation analysis was also conducted to confirm that each indicator provided non-redundant and meaningful contributions. The overall fitness evaluation function is formulated from the combined dataset with 20 effective attributes is presented in Eq. ( 1 ) below: 1 where f 1 is the function having attendance record, number of events participated by the individual student, number of activities of the students per week, attention span and student engagement score. The function f 2 defines the performance of the students in term of their academic profile which contain EP, CD, AD, PD and skill proficiency score. Moreover, the health indicator is defined by the function f 3 that have data of body mass index, cardiovascular scale, health risk and mobile phone activities. Similarly, the function f k defines the instructor effectiveness, student satisfaction, resource allocation, curricula effectiveness and facilities available in the school. A composite vector of sub-functions representing academic success, health markers, and institutional factors makes up the fitness function in Eq. ( 1 ). Specifically, measures engagement metrics like attendance and participation; measures educational achievements, including skill proficiency; measures physiological data like BMI and cardiovascular scale. The weight vector assigns significance factors to each of the 20 qualities in the dataset, regulating their contribution to overall fitness rating. The one of the candidate solutions with its equality constraint required for the optimization of Eq. ( 1 ) using IDA is given in Eq. ( 2 ) 2 where defines the chromosome that is the weight factor for Eq. ( 1 ) with n number of genes, moreover, is the first gene of the chromosome that would contain the optimize value of the first column of the dataset. The weights is not static therefore dynamic optimization of weight vector is achieved via a hybrid IDA. GA generates a diversified population of candidate weight vectors within set constraints. Based on fitness ratings from Eq. ( 1 ), the GA evolves this population through selection, crossover, and adaptive mutation. A Pattern Search (PS) optimizer receives the best-performing chromosome (weight vector) after convergence or function evaluation limit. The PS uses localized search to fine-tune weights without gradient information. The weight vector converges to an ideal configuration that maximizes model accuracy and reduces fitness error over numerous independent runs with dynamic updating. We need to tune weight factor keeping in view the less and more importance of the parameters available in the sub-function defined in Eq. ( 1 ). Once the optimization of the fitness function is performed using proposed IDA the performance and efficacy of PET will be examined. The statistical performance indicators like mean ( ), standard deviation (STD) and kurtosis 38 used to analyze the reliability and effectiveness of the proposed framework is given below in Eq. 3 , Eq. 4 and Eq. 5 , respectively. 3 4 5 where is each individual data point and m is the number of data points. Proposed IDA The proposed IDA framework exploit hybrid version of GA and PS algorithms for training of design parameters, i.e., weight factor . The generic flow diagrams are provided in Fig. 1 while the pseudo developed is given below. PS is a black-box search method that operates without the need for gradients. The algorithm is based on the theory of positive bases and is most effective in d-dimensional spaces 39 . This approach involves exploring the solution space by evaluating a series of points that represent trial solutions in relation to Eq. 1 . The process transitions from a current point to other localized points by employing specified step sizes, which are adjusted based on the outcomes of the evaluated points. This process persists until a specified convergence criterion is achieved, such as the maximum allowable iterations or the desired improvement value 40 . This study used the PS algorithm as the local optimizer in our hybrid IDA framework because it works well in derivative-free optimization settings with complex, non-convex, and noisy fitness landscapes. PS may efficiently explore the local neighborhood of the solution space without analytical gradients, unlike gradient-based approaches. In low-dimensional environments like our 20-parameter weight vector, PS requires fewer hyperparameter modifications and is less prone to premature convergence than population-based heuristic PSO. Bayesian Optimization is good for low-dimensional, expensive function evaluations, but its probabilistic surrogate models may not scale well across 100 simulation runs or dynamic fitness landscapes. Our ablation results show that PS’s deterministic and computationally consistent search method complements GA’s exploratory nature, resulting in faster local convergence and improved fitness evaluation stability. PS was chosen for hybrid optimization because it is balanced, computationally efficient, and robust. The PS utilized in this study incorporates input from one of the optimal outcomes of GA, a global optimizer, and subsequently conducts a local exploration of the best-optimized parameters to enhance fitness, convergence rate, and stability 41 , 42 . GA, introduced by Holland in 1975, are part of the evolutionary computation family and are inspired by natural evolutionary processes. The optimization fitness of the global optimizer relies on the diversity of the initial population and the reproduction mechanism. The boundary of the search space in the genetic algorithm is regulated using a log-sigmoidal function, and the optimal solution is achieved by appropriately selecting the fittest chromosomes in each generation. The bounded genetic algorithm (GA) has been utilized on the fitness function, considering the established advantages of GA, including its ability to prevent premature local convergence, maintain search diversity, and ensure computational efficiency. It has been utilized for parametric optimization 43 and integrated with artificial neural network modeling across various linear and nonlinear systems in social sciences, engineering, and other fields 31 , 44 , 45 . The pseudo code developed for the IDA scheme is presented below: MaxFunEvals: Predefine number of maximum fitness evaluations, i.e., 350,000 FitEvals: Predefine number of fitness evaluations, i.e., FitEvals = MaxFunEvals/2=175,000 Start: GA // start genetic algorithm Inputs: Size of population, i.e., alpha Rate of elitism, i.e., betaOutput: Candidate solution of GA BeginDecoded chromosomes as a feasible candidate Set of chromosomes, i.e., initial population, Papu. Set values for Fitness → 10 -12 , TolFun → 10 -15 , TolCon → 10 -12 , TimeLimit → ∞, StallTimeLimit → ∞ and StallGenLimit → ∞ such that GA termination criteria never fulfilled expect FunEvals Set number of evaluations, nEvals → 0 While { nEvals is less than or equal to FitEvals } do % FitEvals = MaxFitEvals/2Evaluate the fitness individual of population using relation in Eq. ( 1 )Repeat for α of Papu nEvals = nEval + 1 Rank each Individual with respect to the minimum of their fitness evaluation Function call ‘@selectionstochastiuniform’ Function call ‘@crossoverheuristic’ Function call ‘@mutationadaptfeasible’ Migration interval equal to 30 Update the Papu and proceed the algorithm form fitness evaluation step Store the optimistic candidate solution , and computational budget as “T”End GA // end genetic algorithm Start PS // Pattern search startBeginInitialize PS algorithm with global best of GAs Set the values of Fitness → 10 -12 , TolFun → 0, TolCon → 0 and TolX → 0, such that algorithm never stop for these termination criteria. Set the values of number of evaluations, i.e., nEvals → 0, While { nEvals is less than or equal to FitEvals } do Evaluate the fitness of the using Eq. ( 1 ) nEvals = nEvals +1 Function call ‘@patternsearch’ Update for each cycle of PS procedure. Go to fitness calculation step The final weights of PS along with its fitness and time consumeEnd PS // End Patter Search Repeat the complete procedure for 100 runs to generate large data set for reliable statistics. To avoid data leakage, all hyperparameter tuning was conducted strictly within each training fold of the cross-validation process. For every iteration, the training subset was internally used for grid search, while the corresponding validation fold remained completely unseen during tuning. The final evaluation was carried out exclusively on the held-out test fold, ensuring that no information from the test data influenced model selection or optimization. To ensure robust evaluation, a stratified 5-fold cross-validation scheme was employed. The dataset was randomly partitioned into five equal-sized folds, maintaining the statistical distribution of target variables across each subset. For every fold, the model was trained on four parts and tested on the remaining one. This process was repeated five times so that each subset was used as a test set once. The reported results such as accuracy, performance scores, and computational time are averaged over these five runs. This consistent evaluation protocol was applied across all models including IDA, Random Forest, ANFIS, and SVM to ensure unbiased and reliable comparison. To ensure robust model evaluation, a stratified 10-fold cross-validation strategy was employed. The dataset was split into 10 equally sized folds, preserving class distributions. For each iteration, the model was trained on 90% of the data (9-folds) and tested on the remaining 10% (1-fold). This process was repeated 10 times so that every sample served as a test instance exactly once. Final results were computed by averaging performance across all folds. No additional holdout set was used. Hyperparameter optimization was performed within each training fold using grid search. The parameter spaces were: Random Forest: n_estimators = [50, 100, 150, 200]; max_depth = [5, 10, 15, 20]; min_samples_split = [2, 5, 10]. SVM (RBF kernel): C = [0.1, 1, 10, 100]; gamma = [0.001, 0.01, 0.1, 1]. ANFIS: membership functions = [2, 3, 4, 5]; types = [‘gaussmf’, ‘trimf’]; training epochs = [50, 100, 150, 200]. All tuning was conducted exclusively on the training portion of the cross-validation split to prevent information leakage. Results and discussion The proposed intelligent framework is validated on the private developed dataset as presented in Table 4 . The results are also compared with random forest method 6 , RFR technique 8 SVM, NB and GRNN method 6 . The simulations are performed using MATLAB R2023b on the hardware with the specifications Intel (R) Core processor, 64 GB of RAM, 256 SSD with an integrated 1 TB HDD, and an 8 GB NVIDIA RTX graphics card. The generic as well as specific parameter values and setting used for the optimization of the weight factors of the fitness function developed in Eq. ( 1 ) is given in Table 2 . These specific parameters value of GA and PS tabulated in table are settled after a cumbersome process by keeping in view the overfitting and underfitting of the solvers. Similarly, papulation size, crossover rate, mutation rate etc. are established keeping in view different parameters like convergence rate, fitness value, results approximation and others. Moreover, it is also worth to mention that number of generations are adjusted by observing the rate of fitness improvement over the number of generations. Table 4. Performance of proposed IDA technique with change in learning rate. Learning rate Performance indicator Value OP EP HI 0.0135 ✓ 95.3 ✓ 82.2 ✓ 0.81 0.0013 ✓ 96.5 ✓ 82.9 ✓ 0.82 0.00046 ✓ 97.1 ✓ 84.7 ✓ 0.82 0.00013 ✓ 98.0 ✓ 85.9 ✓ 0.84 Open in a new tab Table 2. Parameter values and setting for IDA. Parameters Value/settings Number of variables 20 Lb Lower bound of variables Ub Upper bound of variables Creation function Constraint dependent Selection function Stochastic uniform Scaling function Rank Elite count 0.15 Crossover fraction 0.85 Crossover function Heuristic Mutation Adaptive feasible Migration Both Migration fraction 0.3 Penalty factor 75 Stall time 50 Number of generations 200 Function tolerance 1×e − 9 Options Default Hybrid Function Pattern Search Open in a new tab A hybrid intelligent approach is employed to extract an optimized weight vector, aiming to maximize the fitness of the proposed techniques and enhance the performance of college students in terms of education, mental health, and physical health. The fitness function is optimized by integrating a pattern search algorithm as a local optimizer with the parameters derived from the genetic algorithm. The students’ performance is evaluated in relation to the established baseline of their performance prior to the implementation of the physical education programs. To evaluate fairness, the proposed framework was tested separately on male and female subsets. The model demonstrated comparable performance in both groups, with no statistically significant difference in overall composite evaluation scores. However, slightly higher variance was observed in the female subgroup due to its smaller representation. These results emphasize the importance of demographic balance in datasets and support the need for further research into fairness-aware optimization methods. The training utilized the proposed IDA technique over 200 generations, resulting in a mean fitness value of 98.102% upon execution of the PET, as illustrated in Fig. 3 . The figure indicates that the average distance between individuals varies across generations, ultimately converging to yield optimized results in intelligent algorithms. The fitness scaling depicted in the figure illustrates the variability in expectation values across generations. Fig. 3. Open in a new tab Mean and Best fitness achieved after PET implementation for college students. The fitness function, or objective function, is effectively optimized when all adaptive weighted values are fine-tuned. As illustrated in Fig. 4 , the fitness of each individual is consistently observed at 98, with scores ranging from 98 to 98.4. The genealogical profile graph presented in the figure compares the genetic profiles of individuals who submitted their data to identify potential familial relationships within the dataset. Fig. 4. Open in a new tab Fitness of each individual, histogram score and genealogical profile. The number of offsprings generated for 2000 individual during the selection process of the offspring from one generation to the other generation is presented in Fig. 5 along with the percentage of the stopping criteria met once the optimized weight vector is obtained of the length of 20 adaptive parameters. Fig. 5. Open in a new tab Selection process of the offspring and percentage of the stopping criteria met. Once the IDA technique is executed, the weight factor in the form of a vector is exported and performance of the students is determined using different performance indicators like overall performance (OP), educational performance (EP), health indicator (HI), cognitive domain (CD), affective domain (AD) and psychomotor domain (PD). One thing is important to mentioned that the mentioned indicators are average results from the complete dataset. The same parameters have also been evaluated before the execution of PET programs. The percentage of improvement in all the indicator after the execution of PET program is tabulated in Table 3 . Table 3. Average students educational and personal improvement due to physical educational trainings. Students’ performance indicators Before PET After PET Improvement (%) OP 39.0 98.0 151.128 EP 71.8 85.9 19.6370 HI 0.65 0.84 29.2300 CD 70.0 82.0 17.1428 AD 65.0 78.0 20.0000 PD 58.0 73.5 26.7241 Open in a new tab As shown in Table 3 , the students’ educational performance improved by 19.64% following the implementation of the PET program. Moreover, the affective domain and psychomotor domain has been improved in the students by 20% and 26.7241%, respectively that is better than the cognitive domain. This represents that the physical educational activities among the college students instigates the eager of competition and sportsman ship among the students that increases their attention for the verbal communication, presentation skills and laboratory skills. It is also worth mentioning that the health indicator has also been increased from 0.65 to 0.84 that generates good energy among the students and their emotional activities. The analysis of the best, worst and means scores for 200 generations are also plotted in Fig. 6 and results are drawn after each 10 generation of IDA proposed hybrid approach. The overall performance of the college students before PET program is drawn in Fig. 6 (a) while the results after PET program implementation are presented in Fig. 6 (b). From the figure it is quite clear that overall performance of the students lies from 38 to 40 before the PET applied among the students that is quite improved once the program has been implemented. Fig. 6. Open in a new tab After physical education training. ( a ) Overall performance before PET. ( b ) Overall performance after PET. Figure 6 (a) illustrates the overall performance of college students before the implementation of the PET program. The y-axis represents the performance scores, while the x-axis denotes the number of generations over which the students’ baseline scores were simulated using the proposed IDA framework. Figure 6 (b) presents the overall performance after the PET program, showing a significant improvement across generations. The increase in performance indicates the effectiveness of the PET program in enhancing cognitive, affective, and psychomotor development. The gap between the two curves (before and after PET) visually highlights the impact of the optimized training strategy. Both subfigures include best, worst, and mean performance scores drawn at regular 10-generation intervals over 200 generations. These results demonstrate consistent improvement and convergence behavior enabled by the IDA optimization. Ablation study-1: Convergence of the proposed algorithm The convergence of the proposed algorithm is dependent on the learning rate that also affects the accuracy as well as the computational complexity in term of time and space. In this regards performance of the students is examined for various values of the learning rate and best results are observed for the learning rate of 0.00013 and relatively lesser for 0.0135 while rest of the parameter values and setting of the IDA algorithm are kept constant and results are tabulated in Table 4 . It is clear from the table that the values of the performance indicator like OP, EP and HI increases with the decrease in learning rate. It is worth mentioning that the computational budget will also increase but this effect can be mitigated keeping in view the availability of sophisticated hardware. Ablation study-2: Reliability of the proposed IDA The reliability of the proposed IDA is developed by using sufficient number of runs executed in an independent manner. In this regard 100 independent runs are executed and in each independent run the number of generations were the same as 200 and the results are plotted in Fig. 7 . Fig. 7. Open in a new tab Reliability of proposed IDA approach for 100 independent runs. The graph is drawn on the semiology scale to see the clear effects of the error of the fitness function. The error profile for OP, EP and HI is observed to be from the range 10 − 08 to 10 − 06 , 10 − 07 to 10 − 05 and 10 − 06 to 10 − 05 , respectively that ensures the reliability of the proposed intelligent hybrid scheme. Given that the dataset comprises 200 male and 60 female students, concerns of gender imbalance were carefully addressed. A stratified 5-fold cross-validation strategy was used to ensure proportional gender representation in both training and test sets. Additionally, performance metrics were evaluated separately for male and female students. The proposed IDA demonstrated consistent accuracy across both groups, with only marginal differences that fall within standard deviation bounds. However, we acknowledge the potential benefit of bias mitigation strategies such as sample reweighting, gender-balanced resampling, or adversarial debiasing. These enhancements are planned for future research to further strengthen fairness and generalizability in gender-sensitive educational assessments. Ablation study-3: Overall performance and stability of the model on 100 independent runs The single good result in stochastic based intelligent schemes is not good enough to ensure the performance of the intelligent algorithm. In this regards the global fitness values and global values of the OP, EP and HI are also tabulated in Table 5 . The global fitness and performances are calculated upto three decimal places and the basic statistical indicators like minimum (Min), maximum (Max), mean (Mean), standard deviation (STD) and kurtosis are also calculated to see the supremacy of the proposed approach. Keeping in view the mean and STD values of the fitness and performance of 100 independent runs the stability of the proposed IDA is guaranteed. Table 5. Global performance and stability of proposed IDA. Statistical indicators Global fitness Values Global performance Values OP EP HI OP EP HI Min ✓ 1.3028E-08 ✓ 97.89 ✓ 1.0908E-07 ✓ 82 ✓ 1.0812E-06 ✓ 0.79 Max ✓ 1.0670E-07 ✓ 98.5 ✓ 9.9759E-06 ✓ 85.3 ✓ 1.0947E-05 ✓ 0.84 Mean ✓ 6.4309E-08 ✓ 98.0 ✓ 5.0852E-06 ✓ 85 ✓ 5.4689E-06 ✓ 0.86 STD ✓ 2.7669E-08 ✓ 1.6532 ✓ 2.9793E-06 ✓ 1.6734 ✓ 2.8767E-06 ✓ 1.7453 Kurtosis ✓ 1.7238 ✓ 1.6452 ✓ 1.7282 ✓ 1.8358 ✓ 1.8218 ✓ 1.8340 Open in a new tab Ablation study-4: Computational budget in term of time The physics of the proposed algorithm is heuristic like other artificial intelligence based computational techniques therefore its computational time cannot be found like other deterministic algorithms. In this regard, the computational complexity in term of time in seconds is calculated for 100 independent runs on the machine whose hardware is mentioned in the start of simulation and results section. The graph is drawn in Fig. 8 to see the complete time profile of each independent run. Fig. 8. Open in a new tab Computational time in seconds for 100 independent runs. From the figure, the minimum time was observed to be 1201 s, while the maximum time taken by one of the independent runs was found to be 1500 s. Realistically speaking the mean computational time, standard deviation in the time of 100 runs and its kurtosis values was calculated as 1353.48 s, 91.012 s and 1.776 s, respectively. Comparison with state of art and discussion The results of the proposed IDA in term of the mean execution time (MET) and performance are presented in. Table 6 . It is quite clear from the table that the proposed IDA achieved the highest performance at 98%, demonstrating its superiority in this context. In contrast, the RF method yielded an 88.21% performance level, which, while still effective, falls significantly short of the IDA’s results. The proposed IDA has a deficiency of only 2%, indicating minimal room for improvement, while the RF method shows an 11.79% deficiency, the same sort of trend is observed with other state pf art and reported techniques. The MET value provides insight into the workload associated with each method. The proposed IDA has a MET of 1353.5, which is relatively efficient compared to other methods. The RF method has a MET of 1486, suggesting it may require more resources or exert more effort relative to the performance it provides. Table 6. Comparison of the proposed IDA with reported state of art techniques. Methods Performance measure Performance (%) Deficiency (%) MET Proposed IDA 98 2 1353.5 Random forest 6 88.21 11.79 1486 ANFIS network 7 94.49 5.51 2039 RFR 35 82 18 1529 SVM (kernal = rbf) 84.64 15.36 1374 GRNN Method 6 84.36 15.64 1367 NB 6 84.35 15.65 1359.5 Open in a new tab To statistically validate the performance improvements of the proposed IDA framework, we conducted both Wilcoxon Signed-Rank and paired t-tests comparing its results with those of RF, ANFIS, SVM, GRNN, and NB across 5-fold cross-validation. In all cases, the p-values were less than 0.01, confirming that the observed gains are statistically significant and not due to random variation. Furthermore, we expanded our baseline comparison by including recent hybrid models referenced in Section II, ensuring a more exhaustive and unbiased benchmark evaluation. As compared with the reported results the proposed architecture due to its hybrid nature has the ability to adapt to data complexity, capture non-linear multi-domain relationships, and optimize feature contributions more effectively than conventional models. These additions provide a stronger theoretical and empirical justification for the performance gains observed in Table 6 . As part of our benchmarking, classical ML models including NB, SVM, and RF were evaluated using the same cross-validation and performance metrics. Results demonstrate that the proposed IDA consistently outperforms these models in both accuracy and robustness as shown in Table 6 . To ensure a fair comparison between the proposed IDA and existing models, we employed a consistent evaluation framework using identical data splits and cross-validation settings. Hyperparameters for the RF, ANFIS, and SVM models were tuned using a grid search approach. For RF, we varied the number of estimators (ranging from 50 to 200), maximum depth (5 to 20), and minimum samples split (2 to 10). For ANFIS, membership function types (e.g., gaussmf, trimf), number of membership functions (2 to 5), and training epochs (50 to 200) were considered. For SVM, we explored different kernel types (linear, polynomial, RBF), regularization parameters (C values from 0.1 to 100), and kernel coefficients (gamma values from 0.001 to 1). The best configurations were selected based on cross-validation accuracy. This ensured that all models were optimized to their best potential within the same experimental setup, thereby supporting a fair and unbiased performance comparison. To assess the computational efficiency of the proposed IDA relative to baseline models, a detailed cost analysis was conducted, considering both training and inference times. Table 7 presents the training duration (in seconds) and average inference time per sample (in milliseconds) across all evaluated models. The results indicate that while the IDA demands a longer training time due to its iterative evolutionary and local search procedures it achieves a comparable inference efficiency. Notably, IDA’s inference time is marginally faster than ANFIS and nearly on par with RF and SVM. Given its superior accuracy, the computational trade-off during training is justifiable for offline and decision-support applications in educational domains. Table 7. Computational cost comparison of evaluated models. Model Training time (s) Avg. inference time (s) Proposed IDA 1353.5 5.8 RF 186.2 5.3 RBF Kernal 221.4 5.6 ANFIS Network 442.7 6.4 Open in a new tab While this study focused on comparing the proposed hybrid intelligent decision algorithm with widely-used machine learning models such as RF, ANFIS, and SVM, we acknowledge the growing relevance of deep learning techniques in performance modeling. Models like MLPs and Transformer-based architectures have demonstrated exceptional capabilities in pattern recognition and sequence modeling. However, their application was deferred in the present work due to the limited sample size of the dataset and the need for interpretability and computational feasibility. Future research will aim to incorporate and compare such deep learning frameworks, particularly as larger and more diverse datasets become available, to further strengthen the evaluation and generalization of intelligent physical education assessment systems. For qualitative interpretation, student performance improvement was categorized as follows: positive (> 10%), neutral (0–10%), and negative (< 0%). The distribution across the full dataset was: positive (78%), neutral (16%), and negative (6%). These categories helped visualize the real-world impact of the training program and are reported as supplementary analysis in Table 8 . Table 8. Computational cost comparison of evaluated models. Category Range % of students Positive Improvement > 10 78 Neutral Change 0–10 16 Negative Change < 0 6 Open in a new tab To complement the quantitative results, we gathered informal qualitative feedback from ten students and four physical education instructors involved in the private dataset institution. After reviewing the system-generated outputs and visual analytics, instructors reported that the IDA’s ability to distinguish between academic, affective, and psychomotor domains enabled more informed and personalized training decisions. They particularly valued the ability to identify students needing focused support. Students indicated that visualizing their progress over time helped them better understand the purpose of physical education activities and motivated further participation. These early insights suggest the system is not only technically effective but also practically acceptable and educationally meaningful in real-world classroom settings. The Wilcoxon Signed-Rank Test and Paired t-test results from two statistical significance tests comparing the proposed IDA approach to other methods are shown in Table 9 . The Proposed IDA had the highest statistically significant findings, with Wilcoxon and t-test p-values of 0.0029 and 0.0016, respectively, far better than baseline models. Random Forest (RF) and RFR similarly had t-test p-values below 0.05 (0.0473 and 0.0375), but their Wilcoxon p-values were above the threshold, suggesting their performance differences are less robust. ANFIS network showed statistical significance in Wilcoxon test (0.0373) but not paired t-test (0.0626), indicating inconsistent significance. SVM (RBF kernel), GRNN, and Naive Bayes had p-values above 0.05 in both tests and did not vary significantly. Both non-parametric and parametric testing show that the suggested IDA method outperforms previous methods. Table 9. Statistical tests for the proposed IDA and other reported techniques. Methods Statistical test Wilcoxon Signed-Rank Test Paired t-test Proposed IDA 0.0029 0.0016 RF 0.0591 0.0473 ANFIS network 0.0373 0.0626 RFR 0.0636 0.0375 SVM (kernal = rbf) 0.0672 0.0484 GRNN Method 0.0655 0.0594 NB 0.0694 0.0609 Open in a new tab Model adoptability and deployment in diverse educational contexts The practical utility of the proposed IDA extends beyond the confines of a single institutional or cultural context. Although the present study was conducted using data from one educational institution, the framework was intentionally designed to be modular, data-driven, and generalizable to support deployment across diverse academic, demographic, and cultural settings. Academic versatility The IDA can be customized to align with varying curricular goals, including cognitive, psychomotor, and affective learning objectives defined by national education standards or institutional strategies. Performance indicators used in the fitness function such as academic progress, health metrics, or behavioral engagement can be adjusted based on the local assessment structure or educational stage. The underlying weight optimization mechanism dynamically adapts to these indicators, making the model applicable in both standardized and competency-based education systems. Demographic adaptability The model can accommodate demographic diversity by integrating personalized features such as age, gender, socioeconomic status, and digital access. For instance, in contexts where mobile phone usage or attendance patterns significantly vary by demographic segment, the model adjusts weight distributions through evolutionary optimization, ensuring equitable evaluation. Furthermore, the stratified cross-validation protocol ensures that subgroup representation is preserved during model training and evaluation. Cultural and regional transferability Cultural relevance is preserved by allowing local educators to define or prioritize custom performance metrics reflecting community values. In international or multilingual settings, subjective inputs such as student satisfaction or instructor feedback can be collected using localized instruments and interpreted through culturally responsive rubrics. The empirical weight learning mechanism ensures that even when assessment dimensions are redefined or reweighted, the optimization remains effective without manual recalibration. Implementation guidelines To deploy IDA across multiple sites, the following steps are recommended: Data Collection: Use standardized digital forms to collect academic, physical, and behavioral records tailored to local objectives. Attribute Mapping: Align local indicators with the model’s expected input schema. Local Calibration: Perform initial model runs using site-specific data to tune the weight vector. Interpretability Layer: Provide dashboard visualizations or narrative summaries for educators and students to ensure transparency and usability. Future directions for multi-site evaluation To ensure robustness and global applicability, future research should prioritize multi-institutional and cross-cultural evaluations by deploying educational models across diverse schools and universities in both urban and rural settings. This approach would help uncover the sensitivity of educational technologies to environmental, pedagogical, and cultural variations. Key factors to consider include implicit knowledge development in EFL learners 46 , deep learning strategies in SPOC environments 47 , and the comparative sustainability of global online education platforms 48 . Further exploration should involve predictive models for cognitive performance 49 , explainable knowledge tracing systems 50 , and innovative platform designs such as those based on the meta-universe 51 . Public service platforms and national smart education strategies also warrant evaluation under real-world policy frameworks like China’s “Double Reduction” policy 52 . Social support mechanisms in digital learning communities 12 , the effectiveness of media-enhanced strategies in specialized fields like pharmacology 53 , and the unique challenges of rural education systems 54 are additional dimensions that can enrich future multi-site research and guide the design of more inclusive and adaptive smart education systems. Limitations While the proposed IDA demonstrates high performance and robustness across multiple indicators, certain limitations must be acknowledged to provide a balanced perspective on its applicability and generalizability. Risk of overfitting Although the model was trained and evaluated using stratified 10-fold cross-validation, the integration of evolutionary optimization (GA) and local search (PS) inherently involves high model capacity and tuning flexibility. This increases the risk of overfitting, particularly when working with moderate-sized datasets or highly specific performance metrics. While repeated runs, statistical evaluations, and regularization constraints were used to mitigate this risk, there remains a possibility that the model could capture noise or local patterns that do not generalize well to unseen student populations. Limited geographic and institutional diversity The dataset used for training and validation includes students from a single institutional and regional context, primarily reflecting urban, undergraduate-level participants within a specific country. As a result, the model may not fully capture the variance introduced by cross-regional, cross-institutional, or cross-cultural educational environments. Factors such as language, assessment styles, curriculum design, and learning behaviors may differ significantly in other regions, affecting model performance when transferred directly. These constraints underline the importance of future validation studies across multi-site, rural-urban, and international cohorts to ensure the model’s adaptability and fairness. Enhancing the dataset diversity and incorporating domain-specific regularization or fairness constraints will also be essential to reduce overfitting and bias in broader deployments. Implementation requirements and institutional readiness For effective deployment of the IDA in educational institutions, both technical infrastructure and human capacity must be adequately prepared. This section outlines the practical steps and foundational requirements to support successful model adoption. Infrastructure requirements The IDA framework is lightweight in deployment and can operate on standard computing infrastructure. However, the following components are recommended to ensure smooth integration: Data Management System: Institutions should maintain a centralized digital repository that includes academic records, attendance logs, health screening data, and survey feedback. These data sources form the input attributes for the model. Computing Resources: The optimization and model training can be executed on a standard workstation with modern CPU and at least 16GB RAM. For institutions with higher student populations or concurrent processing needs, a server or cloud-based setup with GPU acceleration (e.g., via AWS, Azure, or institutional HPC clusters) can be employed. Software Stack: The model is implemented using MATLAB (or can be translated into Python), requiring installation of relevant toolboxes or open-source packages for evolutionary computation, optimization, and visualization. Staff training and capacity building The success of the IDA framework depends on its interpretability and usability by non-technical educational staff, including physical education instructors, counselors, and academic coordinators. Institutions are encouraged to: Train Decision-Making Staff: Conduct orientation workshops for educators on understanding performance indicators, model output interpretation (e.g., weights, performance trajectories), and intervention planning. Develop User Interfaces: Provide staff with user-friendly dashboards or data visualization tools that convert model outputs into actionable insights without needing direct algorithmic interaction. Form Interdisciplinary Teams: Implementation should be overseen by a team involving IT personnel, data analysts, physical educators, and administrators to ensure alignment between technical setup and educational goals. Data governance and privacy Institutions must establish protocols for data privacy, anonymization, and ethical use of student information. Consent procedures and compliance with national or regional data protection regulations (e.g., GDPR, FERPA) should be incorporated into deployment planning. Pilot and phased rollout It is advisable to begin with a pilot phase involving a subset of classes or departments to monitor usability and system response. Feedback collected during this phase should inform larger-scale deployment and adaptation. Conclusions Based on the comprehensive results supported by the accompanying graphs and tables, the conclusions are drawn, in this regard the training efficacy and student outcomes have significantly improved, alongside enhancements in health indicators. The proposed framework demonstrates superiority over existing techniques such as RF, ANSI networks, RF regression models, GRNN methods, NB, and SVM utilizing radial basis kernels, achieving a mean performance of 98%. The proposed IDA ensures convergence across various learning rates, with optimal results achieved at a learning rate of 0.00013. The fitness values for overall performance, educational performance, and health indicator score are found to range from 10 − 07 to 10 − 08 , 10 − 07 to 10 − 06 , and 10 − 07 to 10 − 05 , respectively. The findings presented in Figs. 7 and 8 , and Table 6 , based on 100 independent runs, demonstrate the applicability, reliability, and stability of the proposed IDA. Future research may explore the fuzzy structure for non-linear problems in conjunction with advanced AI-based algorithms. By embedding core educational theories such as Bloom’s Taxonomy and Constructivist Learning Theory into the structure of our framework, we establish a pedagogically sound foundation for the integration of artificial intelligence in physical education program assessment. One key limitation of this study is the lack of demographic diversity in the student sample. The data were collected from a single country and institution, which may not capture the variability in educational practices and student behaviors across different cultural or socioeconomic contexts. Future research should aim to validate the proposed framework using multi-site datasets that include rural and urban populations, diverse educational systems, and international cohorts. Cross-cultural validation would help evaluate the adaptability and fairness of the IDA in broader applications. Future work will focus on testing the model on datasets from institutions in different regions or educational systems to assess generalizability and transferability of results. In the future, the authors will extend the same study for state-of-the-art deep learning models like MLPs or Transformers. Additionally, future research may benefit from incorporating advanced methodologies, as demonstrated in works like 55 – 59 , which explore innovative machine learning and AI frameworks for educational and industrial applications. Specifically, these works 60 – 62 focus on optimizations in federated learning, resilience, and cognitive ergonomics, offering pathways to enhance the proposed framework’s robustness and adaptability. Author contributions Xinlong Wang: Conceptualization, methodology, software development, data analysis, writing—original draft preparation.Wenze Sun: Supervision, methodology, writing—review and editing, project administration.Both authors Contributed to the study design, manuscript revisions, and approval of the final version. Data availability The dataset used in this study consists of a publicly available Kaggle dataset and a private dataset obtained from a local college. A publicly available Kaggle dataset and the code used in this study is now publicly available and can be accessed via the following GitHub link: https://github.com/knowledge4every1/Students-Health-and-Academic. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. of Health, U. S. D. & Services, H. The association between school-based physical activity, including physical education, and academic performance, (2010). [ DOI ] [ PubMed ] 2. Marques, A., Corrales, F. R. G., Martins, J., Catunda, R. & Sarmento, H. Association between physical education, school-based physical activity, and academic performance: a systematic review, Retos: nuevas tendencias en educación física, deporte y recreación , vol. 31, pp. 316–320, (2017). 3. Palao, J. M., Hastie, P. A., Cruz, P. G. & Ortega, E. The impact of video technology on student performance in physical education. Technol. Pedagogy Educ. 24 (1), 51–63 (2015). [ Google Scholar ] 4. Jiao, C., Qian, K. & Zhu, D. Application of flipped classroom teaching method based on VR technology in physical education and health care teaching. IEEE Access , (2020). 5. Lindberg, R., Seo, J. & Laine, T. H. Enhancing physical education with exergames and wearable technology. IEEE Trans. Learn. Technol. 9 (4), 328–341 (2016). [ Google Scholar ] 6. Xu, Q. & Yin, J. Application of random forest algorithm in physical education, Sci. Program. , vol. no. 1, p. 1996904, 2021. (2021). 7. Mojsilović, M. et al. Statistical evaluation of the achievements of professional students by combination of the random forest algorithm and the ANFIS method. Heliyon 9 (11), e11451 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Doz, D., Cotič, M. & Felda, D. Random forest regression in predicting students’ achievements and fuzzy grades. Mathematics 11 (19), 4129 (2023). [ Google Scholar ] 9. Ma, Y., Bennett, D. & Chen, S. Perceived organisational support and university students’ career exploration: the mediation role of career adaptability. High. Educ. Res. \& Dev. 42 (4), 903–919. 10.1080/07294360.2022.2115983 (2023). [ Google Scholar ] 10. Wang, T. et al. TASTA: text-assisted spatial and temporal attention network for video question answering. Adv. Intell. Syst. 5 (4), 2200131 (2023). [ Google Scholar ] 11. Wang, T. et al. ResLNet: deep residual LSTM network with longer input for action recognition. Front. Comput. Sci. 16 (6), 166334. 10.1007/s11704-021-0236-9 (2022). [ Google Scholar ] 12. Huang, C. et al. How does social support detected automatically in discussion forums relate to online learning burnout? The moderating role of students’ self-regulated learning. Comput. Educ. 227 , 105213 (2025). [ Google Scholar ] 13. Wu, X., Li, L., Tao, X., Yuan, J. & Xie, H. Towards the Explanation Consistency of Citizen Groups in Happiness Prediction via Factor Decorrelation. IEEE Trans. Emerg. Top. Comput. Intell. 9 (2), 1392–1405. 10.1109/TETCI.2025.3537918 (2025). [ Google Scholar ] 14. Yang, Y., Liu, J., Meng, W., Zhuo, Y. & Wu, Z. An integrated consensus reaching process for product appearance design decision-making: Combining trust and empathy relationships. Adv. Eng. Inform. 67 , 103562. 10.1016/j.aei.2025.103562 (2025). [ Google Scholar ] 15. Liu, Q. et al. CoRLHF: Reinforcement learning from human feedback with cooperative policy-reward optimization for LLMs. Expert Syst. Appl. 301 , 130113. 10.1016/j.eswa.2025.130113 (2026). [ Google Scholar ] 16. Cao, X. et al. Analytical Survey of Learning with Low-Resource Data: From Analysis to Investigation. ACM Comput. Surv. 10.1145/3773075 (2025). [ Google Scholar ] 17. Sajid, M. et al. Advanced multilayer security framework: integrating AES and LSB for enhanced data protection. J. Supercomput . 81 (17), 1607. 10.1007/s11227-022-04590-5 (2025). [ Google Scholar ] 18. Ahmed, A. et al. Efficient Melanoma Detection Using Pixel Intensity-Based Masking and Intensity‐Weighted Binary Cross‐Entropy. Int. J. Imaging Syst. Technol. 35 (5), e70179. 10.1002/ijim.70179 (2025). [ Google Scholar ] 19. Almuhaimeed, A. et al. Brain tumor classification using GAN-augmented data with autoencoders and Swin Transformers, Front. Med. (Lausanne). , vol. 12, p. 1635796, (2025). 10.3389/fmed.2025.1635796 [ DOI ] [ PMC free article ] [ PubMed ] 20. Bilal, A. et al. Explainable AI-driven intelligent system for precision forecasting in cardiovascular disease, Front. Med. (Lausanne). , vol. 12, p. 1596335, (2025). 10.3389/fmed.2025.1596335 [ DOI ] [ PMC free article ] [ PubMed ] 21. Bilal, A. et al. Fuzzy deep learning architecture for cucumber plant disease detection and classification. J. Big Data . 12 (1), 117. 10.1186/s40537-025-00423-w (2025). [ Google Scholar ] 22. Ahmed, A., Sun, G., Bilal, A., Li, Y. & Ebad, S. A. A hybrid deep learning approach for skin lesion segmentation with dual encoders and Channel-Wise attention. IEEE Access. 13 , 42608–42621. 10.1109/ACCESS.2025.4289031 (2025). [ Google Scholar ] 23. Wasim, M., Ahmed, I., Ahmad, J. & Hassan, M. M. A novel deep learning based automated academic activities recognition in cyber-physical systems. IEEE Access. 9 , 63718–63728 (2021). [ Google Scholar ] 24. Trudeau, F. & Shephard, R. J. Physical education, school physical activity, school sports and academic performance. Int. J. Behav. Nutr. Phys. activity . 5 , 1–12 (2008). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Donnelly, J. E. & Lambourne, K. Classroom-based physical activity, cognition, and academic achievement. Prev. Med. (Baltim) . 52 , S36–S42 (2011). [ DOI ] [ PubMed ] [ Google Scholar ] 26. McKenzie, A. I. Functional Effects of Personality on Individual and Team Sport Success ( University of Windsor, 2021). 27. Di Franco, G. & Santurro, M. Machine learning, artificial neural networks and social research. Qual. Quant. 55 (3), 1007–1025 (2021). [ Google Scholar ] 28. Doucette, J., Larson, K. & Cohen, R. Conventional machine learning for social choice, in Proceedings of the AAAI Conference on Artificial Intelligence , (2015). 29. Khan, J. A. et al. Design of Super Resolution and Fuzzy Deep Learning Architecture for the Classification of Land Cover and Landsliding using Aerial Remote Sensing Data. IEEE J. Sel. Top. Appl. Earth Obs Remote Sens , (2024). 30. Kukker, A. & Sharma, R. A genetic algorithm assisted fuzzy Q-learning epileptic seizure classifier. Computers \& Electr. Eng. 92 , 107154 (2021). [ Google Scholar ] 31. Riaz, F. M. et al. Numerical Treatment of Non-Linear System for Latently Infected CD4 + T Cells: A Swarm-Optimized Neural Network Approach. IEEE Access , (2024). 32. Kukker, A. & Sharma, R. JAYA-optimized fuzzy reinforcement learning classifier for COVID-19. IETE J. Res. 69 (10), 7314–7325 (2023). [ Google Scholar ] 33. Bose, I. & Mahapatra, R. K. Business data mining—a machine learning perspective. Inform. \& Manage. 39 (3), 211–225 (2001). [ Google Scholar ] 34. Kukker, A., Sharma, R., Pandey, G. & Faseehuddin, M. Fuzzy lattices assisted EJAYA Q-learning for automated pulmonary diseases classification. Biomed. Phys. Eng. Express . 10 (6), 65001 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 35. Bharadiya, J. P. Machine learning and AI in business intelligence: Trends and opportunities. Int. J. Comput. (IJC) . 48 (1), 123–134 (2023). [ Google Scholar ] 36. Kukker, A., Sharma, R., Mishra, O. & Parashar, D. Epileptic seizure classification using fuzzy lattices and Neural Reinforcement Learning. Comput. Methods Biomech. Biomed. Eng. Imaging Vis. 11 (7), 2290361 (2024). [ Google Scholar ] 37. MFA, I. Students’ Health and Academic Performance, (2023). 38. Zhong, J., Wang, D., Guo, J., Cabrera, D. & Li, C. Theoretical investigations on kurtosis and entropy and their improvements for system health monitoring. IEEE Trans. Instrum. Meas. 70 , 1–10 (2020).33776080 [ Google Scholar ] 39. Yu, W. C. Positive basis and a class of direct search techniques. Scientia Sinica Special Issue Math. 1 (26), 53–68 (1979). [ Google Scholar ] 40. Findler, N. V., Lo, C. & Lo, R. Pattern search for optimization. Math. Comput. Simul. 29 (1), 41–50 (1987). [ Google Scholar ] 41. Raja, M. A. Z., Khan, J. A., Ahmad, S. I. & Qureshi, I. M. Numerical treatment for Painlevé equation i using neural networks and stochastic solvers, in Innovations in Intelligent Machines-3: Contemporary Achievements in Intelligent Systems, Berlin, Heidelberg: Springer Berlin Heidelberg, 103–117. (2013). [ Google Scholar ] 42. Jian, Q. Multimedia teaching quality evaluation system in colleges based on genetic algorithm and social computing approach. IEEE Access. 7 , 183790–183799 (2019). [ Google Scholar ] 43. Zaman, F., Qureshi, I. M., Khan, J. A. & Khan, Z. U. An Application of Artificial Intelligence for the Joint Estimation of Amplitude and Two-Dimensional Direction of Arrival of Far Field Sources Using 2-L-Shape Array, Int. J. Antennas Propag. , vol. no. 1, p. 593247, 2013. (2013). 44. Dony, R. D. & Haykin, S. Neural network approaches to image compression, Proceedings of the IEEE , vol. 83, no. 2, pp. 288–303, (1995). 45. Khan, J. A., Raja, M. A. Z. & Qureshi, I. M. Stochastic computational approach for complex nonlinear ordinary differential equations. Chin. Phys. Lett. 28 (2), 20206 (2011). [ Google Scholar ] 46. Zhang, R. & Chen, Y. What can multi-factors contribute to Chinese EFL learners’ implicit L2 knowledge? International Rev. Appl. Linguistics Lang. Teaching , (2024). 47. Jiang, L., Lv, M., Cheng, M., Chen, X. & Peng, C. Factors affecting deep learning of EFL students in higher vocational colleges under small private online courses-based settings: A grounded theory approach. J. Comput. Assist. Learn. 40 (6), 3098–3110 (2024). [ Google Scholar ] 48. Li, D. & Xing, W. A comparative study on sustainable development of online education platforms at home and abroad since the twenty-first century based on big data analysis. Educ. Inf. Technol. (Dordr) . 10.1007/s10639-025-13400-3 (2025). [ Google Scholar ] 49. Li, D., Tang, N., Chandler, M. & Nanni, E. An optimal approach for predicting cognitive performance in education based on deep learning. Comput. Hum. Behav. 167 , 108607 (2025). [ Google Scholar ] 50. Huang, C. et al. XKT: Toward Explainable Knowledge Tracing Model With Cognitive Learning Theories for Questions of Multiple Knowledge Concepts. IEEE Trans. Knowl. Data Eng. 36 (11), 7308–7325 (2024). [ Google Scholar ] 51. Chen, G., Jin, Y. & Chen, P. Development of a platform for state online education services: Design concept based on meta-universe. Educ. Inf. Technol. (Dordr) . 29 (17), 23605–23629 (2024). [ Google Scholar ] 52. Liu, Y., Cao, S. & Chen, G. Research on the Long-term Mechanism of Using Public Service Platforms in National Smart Education—Based on the Double Reduction Policy. Sage Open , 14 , 1, (2024). 53. Li, Y. et al. Unraveling the effectiveness of new media teaching strategies in pharmacology education under different educational backgrounds: Insights from 6447 students. Eur. J. Pharmacol. 989 , 177255 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 54. Li, B., Li, G. & Luo, J. Latent but not absent: The ‘long tail’ nature of rural special education and its dynamic correction mechanism. PLoS One . 16 (3), e242023 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Xu, G. et al. Event-driven publish/subscribe communication protocol for IIoT by report and traceable ring signature. IEEE Trans. Industr Inf. 21 (9), 6670–6678. 10.1109/TII.2025.3567265 (2025). RAT Ring. [ Google Scholar ] 56. Cao, Z. et al. Understanding the Dimensional Need of Noncontrastive Learning. IEEE Trans. Cybern . 55 (9), 4089–4102. 10.1109/TCYB.2025.3577745 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 57. Li, X. et al. Multitype Game Optimisation: A Two-Stage Fine-Tuning Framework for Multi-Game Optimisation With Large Language Models. CAAI Trans. Intell. Technol. 10.1049/cit2.70108 (2026). [ Google Scholar ] 58. Zhang, Y. et al. Neuroergonomics evaluation of teamwork in a fast-paced communication and shared decision-making task. Int. J. Ind. Ergon. 112 , 103892. 10.1016/j.ergon.2026.103892 (2026). [ Google Scholar ] 59. Xu, H., Chen, M., Wang, M., Lu, J., LLM Agent Use on Teachers’ Curriculum Content Creation. & The Chain Mediating Role of School Support and Teacher Self-Efficacy. Behav. Sci. 16 (1), 124. 10.3390/bs16010124 (2026). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Chen, Y., Li, H., Song, Y. & Zhu, X. Recoding Hybrid Stochastic Numbers for Preventing Bit Width Accumulation and Fault Tolerance. IEEE Trans. Circuits Syst. I Regul. Pap. 72 (3), 1243–1255. 10.1109/TCSI.2024.3492054 (2025). [ Google Scholar ] 61. Xu, G. et al. CBRFL: A framework for Committee-based Byzantine-Resilient Federated Learning. J. Netw. Comput. Appl. 238 , 104165. 10.1016/j.jnca.2025.104165 (2025). [ Google Scholar ] 62. Sun, X., Guo, B., Yang, Y. & Pan, Y. Time for a change! Uprooting users embedded in the status quo from habitual decision-making. Decis. Support Syst. 189 , 114371. 10.1016/j.dss.2024.114371 (2025). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The dataset used in this study consists of a publicly available Kaggle dataset and a private dataset obtained from a local college. A publicly available Kaggle dataset and the code used in this study is now publicly available and can be accessed via the following GitHub link: https://github.com/knowledge4every1/Students-Health-and-Academic. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.6 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• RopeFormer: Cross-Trial Adaptation from Interaction History for Dynamic Rope Manipulation#1049173
• Evaluation of the Brothers United Fatherhood Program#1088701
• Exploring multi-objective optimization scheduling strategies for power dispatch robots based on an improved particle swarm optimization algorithm framework.#354914
• Reinforcement Learning–Based Operational Optimization of a Hybrid Renewable System with Energy Storage#250672
• Deep Reinforcement Learning-Based Multi-Objective Optimization for Energy-Efficient UAV-Mounted STAR-RIS Wireless Networks#231847