Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy 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 Mar 3;16:11953. doi: 10.1038/s41598-026-42082-1 Search in PMC Search in PubMed View in NLM Catalog Add to search An IoT-enabled CRNN framework for secure wearable sensor-based activity recognition in physical education Jun Yuan Jun Yuan 1 College of Physical Education and Health, Changji University, Changji City, 83110 China Find articles by Jun Yuan 1 , YiChao Zhang YiChao Zhang 1 College of Physical Education and Health, Changji University, Changji City, 83110 China Find articles by YiChao Zhang 1, ✉ , Bingjie Chen Bingjie Chen 1 College of Physical Education and Health, Changji University, Changji City, 83110 China Find articles by Bingjie Chen 1 Author information Article notes Copyright and License information 1 College of Physical Education and Health, Changji University, Changji City, 83110 China ✉ Corresponding author. Received 2025 Dec 6; Accepted 2026 Feb 24; 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: PMC13068935 PMID: 41776307 Abstract Recent advancements in digital technologies have profoundly transformed healthcare delivery, sports analytics, and physical activity surveillance. The accelerated proliferation of wearable sensing devices and Internet of Things (IoT) infrastructures has facilitated unprecedented large-scale acquisition of multimodal physiological and kinematic data. However, accurate recognition of complex human activities remains a formidable challenge, primarily due to inadequate modeling of the spatiotemporal dependencies inherent in wearable-sensor signals. To address these fundamental limitations, this paper proposed a novel IoT-oriented activity recognition framework predicated on a Convolutional Recurrent Neural Network (CRNN) architecture, engineered to simultaneously model the spatial and temporal characteristics of multimodal wearable-sensor data streams. The proposed framework synergistically integrates Convolutional Neural Networks (CNNs) for hierarchical spatial feature extraction with Recurrent Neural Networks (RNNs) for temporal sequence modeling, thereby enabling more discriminative and robust activity classification. The methodological pipeline comprises several sequential stages: First, multidimensional wearable-sensor datasets are employed, encompassing physiological and inertial measurements including heart rate variability, triaxial accelerometer readings, gyroscopic angular velocity, magnetometer orientation data, and cutaneous temperature signals acquired from multiple anatomical locations. Second, raw sensor signals undergo preprocessing and temporal segmentation procedures to enhance data quality and optimize temporal representation. Third, spatiotemporal feature representations are learned autonomously within the hierarchical CRNN architecture. Finally, the proposed model is systematically evaluated through a comparative analysis with five representative baseline methods, encompassing both conventional machine learning algorithms and contemporary deep learning approaches. Experimental results demonstrate that the proposed CRNN framework achieves superior performance, achieving 98.2% classification accuracy, 97.2% sensitivity, 99.2% specificity, 97.4% recall, and 97.6% precision on the evaluated wearable-sensor datasets. Compared with existing methodologies, the proposed model consistently achieves higher recognition accuracy and greater generalization robustness, underscoring its efficacy for wearable-sensor-based activity recognition and its considerable potential for broader deployment in IoT-enabled monitoring paradigms and adaptable solutions in educational settings. Keywords: Internet of things, Machine learning, Smart wearables, Physical education, Activity monitoring, Convolutional recurrent neural network (CRNN) Subject terms: Engineering, Health care, Mathematics and computing Introduction In recent years, technological advancements have revolutionized industries worldwide, profoundly impacting healthcare, education, and sports. Advanced data collection devices, including those that leverage IoT, AI, and wearables, have enabled real-time analysis and enhanced user experiences. Altogether, these advances have revolutionized how physical activity is recognized and regulated 1 , 2 . People and organizations commonly use smartwatches and fitness bands to monitor personal and activity-related parameters, such as heart rate, calorie consumption, and steps taken. These tools not only help individuals meet their fitness needs but also support research and data-driven analysis endeavors. Then, the advancement and increasing popularity of IoT have complemented this technological network by enabling devices to work together efficiently, making it easier to track and analyze data 3 , and offering novel ways, and it is becoming crucial in areas such as physical and sports activity monitoring. The adoption of information and communication technologies (ICT) in teaching and training in physical education (PE) has transformed traditional instructional methods 4 . Smart wearables that utilize IoT technologies are widely used in PE to monitor physical performance, activity levels, and customized fitness regimes. These technologies benefit practitioners and analysts, as the results obtained inform decision-making and intervention plans 5 , 6 and identify areas for improvement and quantifiable objectives. Smart wearables also promote diversity by tailoring fitness activities to include individuals who require additional support, ensuring inclusive participation 7 , 8 . However, the actual implementation of these technologies requires a strong supporting base, such as the Internet of Things (IoT), for big data storage and processing. Current developments aim to integrate these tools into robust IoT platforms that enable real-time monitoring and analysis of physical activity data. Despite the progress, challenges such as user acceptance, device cost, and data security need to be addressed to achieve widespread adoption. Machine learning has become a fundamental aspect of contemporary society, enabling advances across numerous fields. The ability to make sense of large and complex datasets has made big data particularly useful in the health, financial, and education sectors. 9 . Common machine learning (ML) algorithms used in data-driven analysis include convolutional neural networks (CNNs) and recurrent neural networks (RNNs), which perform well in both structured and unstructured data. Recent advancements have demonstrated the application of ML in wearable-sensor systems for activity recognition, continuous monitoring, and decision support. For instance, such models can distinguish among activities such as walking, running, and cycling using motion data from wearable devices. Furthermore, these models can be updated based on observed behavior patterns and provide adaptive feedback and recommendations over time 10 , 11 . The integration of IoT technologies further extends the application of ML to real-time data processing and decision-making in complex environments. As hardware capabilities and cloud computing infrastructure continue to evolve, ML-based solutions offer compelling advantages in efficiency, scalability, and applicability across many domains 12 . The integration of machine learning with Internet of Things (IoT)–enabled wearable systems has led to a paradigm shift in physical activity analysis and sports-related research. Physical activity tracking has been significantly enhanced through ML algorithms that analyze data collected from wearable devices. In particular, sensor signals are processed to extract spatial characteristics using convolutional neural networks (CNNs) and temporal patterns using recurrent neural networks (RNNs) to classify activities accurately. This capability is especially relevant for activity analysis, as movement patterns provide contextual information for understanding physical behaviors 13 , 14 . In addition, ML-driven analysis of human motion data enables the identification of performance-related characteristics and potential areas for improvement. Recent studies have shown that ML models can estimate injury risk from abnormal motion patterns or movement trajectories, thereby supporting injury prevention and rehabilitation efforts. Moreover, ML-based approaches enhance engagement in physical activity analysis by enabling adaptive and data-driven feedback mechanisms. Despite these advantages, the application of ML in wearable-based activity analysis also presents challenges, including the need for specialized expertise and concerns related to data protection and privacy. This paper proposes an IoT-supported framework we developed using smart wearables and a Convolutional Recurrent Neural Network (CRNN) architecture to analyze physical activity patterns. In our work, the proposed system integrates IoT-based data acquisition with CRNN-based spatial–temporal analysis to enhance the reliability and effectiveness of human activity recognition. Wearable devices enable continuous data acquisition in our framework, cloud platforms facilitate data transmission and storage, and our CRNN model analyzes spatial and temporal features extracted from wearable-sensor signals. To enable efficient analysis, our system utilizes several key modules: IoT-based data collection, where wearable devices capture multidimensional physiological and motion-related parameters; data preprocessing, which includes cleaning, normalization, and feature extraction to ensure data quality; activity recognition using CRNN, combining CNNs for spatial feature learning and RNNs for temporal sequence modeling; feedback generation through data-driven visualization; and data security, using encryption mechanisms to ensure secure transmission and storage. These modules work together to provide accurate and efficient activity recognition. Firstly, we acquire physiological and inertial signals across multiple dimensions through wearable devices. Secondly, we transmit these data to a centralized server via an IoT gateway for preprocessing. Thirdly, our CRNN model identifies activity classes and temporal patterns based on the processed data. Finally, we present the recognition outcomes to the end user to support performance analysis and further evaluation, rather than direct intervention. Our experimental results demonstrate the feasibility of the system in enhancing the accuracy of activity recognition. Compared with conventional methods, our framework achieves improved performance and scalability. In future work, we will focus on optimizing our framework by integrating additional artificial intelligence models and investigating practical challenges related to computational efficiency, usability, and data security. The key contributions of the paper are as follows: We propose an IoT-oriented framework that integrates smart wearable devices with a Convolutional Recurrent Neural Network (CRNN) to analyze wearable-sensor-based physical activity data, ensuring secure data transmission and efficient activity recognition. We develop a hybrid CRNN model that combines CNNs for spatial feature extraction and RNNs for temporal sequence modeling, enabling more accurate classification of complex physical activity patterns compared with conventional methods. We demonstrate the effectiveness and scalability of our framework through experimental evaluation and data-driven visualization, highlighting its applicability to wearable-sensor-based activity recognition. The rest of the paper is organized by logical order as follows: Section " Literature review " illustrates the literature review of physical activity monitoring; Section " IoT and machine learning applications in physical education " of the work presents the integration of Internet of Things (IoT) technologies and machine learning (ML) in physical education and training is rapidly transforming how physical activities are monitored; Section " Experimental results and analysis " of the paper demonstrates the experimental results and discussion in the domain involving physical activity monitoring for the college student. Finally, the research work is concluded in Section " Conclusion " of the paper. Literature review A burgeoning, rapidly evolving domain within physical activity monitoring and Internet of Things (IoT) research encompasses the application of machine learning (ML) algorithms to wearable sensing infrastructure. The widespread adoption of intelligent wearable devices and IoT technologies has profoundly influenced wearable-sensor-based activity recognition paradigms, facilitating more precise physiological monitoring and sophisticated analytical feedback mechanisms. The proliferation of IoT-enabled wearable devices for tracking physical activity exemplifies a broader trajectory toward data-driven activity analysis, used to examine locomotion patterns, performance metrics, and health-related biosignals. Within this framework, 15 delineate IoT devices equipped with inertial measurement units, such as accelerometers and gyroscopes, capable of capturing wearable-sensor kinematic data, thereby enabling a comprehensive assessment of activity patterns. Similarly, 16 highlighted the pivotal role of multi-sensor wearable systems in improving the accuracy of activity recognition, a critical process for evaluating system performance. Such devices capture both physiological and kinematic parameters (e.g., heart rate, ambulatory step count, and body kinematics), which can be synergistically integrated with ML methodologies to enhance recognition precision. Smartwatches and analogous wearable technologies have emerged as increasingly instrumental in bolstering the reliability of activity monitoring within wearable-sensor ecosystems. From a methodological standpoint, 17 proposed a hybrid model integrating Convolutional Neural Networks (CNNs) and Convolutional Recurrent Neural Networks (CRNNs) for processing wearable sensor data streams, exploiting spatiotemporal structural representations to enhance recognition of dynamic activities, including ambulatory locomotion, running, and cycling. These investigations collectively demonstrate how ML techniques can substantially augment the efficacy of wearable-sensor-based activity recognition systems. Extending beyond seminal studies in wearable-assisted activity recognition, contemporary research has increasingly emphasized adaptive and personalized analytical frameworks. Furthermore, 18 examined the integration of IoT devices with machine learning to develop individualized strategies for physical activity analysis. Their investigation demonstrates that wearable devices can capture activity-related data streams, with the resultant information processed through ML algorithms to facilitate adaptive feedback mechanisms. Nevertheless, several substantive challenges persist 19 in a comprehensive analysis of wearable technology deployment in educational settings, including enduring obstacles to large-scale implementation: privacy concerns related to data governance, storage security vulnerabilities, and the considerable financial costs associated with IoT device procurement. A conspicuous absence of standardization in wearable technologies and their interoperability across heterogeneous platforms continues to impede widespread adoption. Similarly, 20 declared that these impediments can be mitigated through coordinated initiatives that synthesize technological advancement with supportive regulatory frameworks, complemented by collaborative engagement between system architects and application stakeholders. In addition to adaptive methodological approaches, recent scholarly inquiry has concentrated on the technological underpinnings of activity recognition models. From a computational modeling perspective, 21 advocated applying machine learning techniques, particularly Recurrent Neural Networks (RNNs), to temporal sequence analysis of physical activity data, emphasizing their inherent suitability for sequential pattern modeling. Analogously, 22 demonstrated that wearable-sensor data can be effectively leveraged to predict activity intensity levels utilizing ML-based predictive approaches. In a related investigation, 23 explored the deployment of deep learning architectures, particularly CNNs, for identifying movement patterns potentially indicative of physiological fatigue, biomechanical inefficiency, or injury predisposition. Moreover, 24 discussed the convergence of IoT and machine learning technologies, accentuating the criticality of secure data management infrastructures such as blockchain-based distributed ledger systems. Collectively, these studies substantiate that while ML-based approaches have yielded promising results, substantive challenges related to robust spatiotemporal feature modeling, cybersecurity considerations, and cross-domain generalization capacity remain unresolved. The extant literature provides valuable theoretical and empirical insights into IoT- and wearable-based activity recognition, offering increasingly accurate computational solutions. For instance, 15 employed deep reinforcement learning algorithms to optimize resource allocation in IoT-based training networks; however, their framework does not explicitly address activity recognition tasks. Similarly, 20 proposed an IoT-based architectural framework that emphasizes system-level infrastructure support rather than granular activity classification. Other IoT-based systems prioritize system architecture design and handcrafted feature extraction methodologies but do not incorporate advanced hybrid learning architectures such as Convolutional Recurrent Neural Networks (CRNNs), which simultaneously capture spatial feature hierarchies and temporal dependencies. Furthermore, 25 investigated secure data transmission protocols in healthcare IoT applications, concentrating on data confidentiality rather than activity recognition performance. In contrast, CRNN-based approaches explicitly model both spatial feature representations and temporal dynamics inherent in wearable-sensor data streams, thereby enabling superior recognition performance. As summarized in Table 1 , the comparative analysis highlights the limitations of existing methodological approaches and provides compelling justification for the CRNN-based framework proposed in this study. Table 1. Comparative Analysis of Activity Recognition Models. Study Accuracy Methodology Key Features 15 Not reported Deep reinforcement learning for resource management in physical education training networks Focus on resource optimization rather than direct activity recognition 20 Not reported IoT-based intelligent physical support framework for physical education Emphasis on educational leadership and student interaction, not solely on activity recognition 3 Not reported IoT-based system for physical activity recognition in physical education Focus on system design and implementation for activity recognition 25 Not reported Survey of IoT and monitoring systems for healthcare applications Discussion on secure transmission and management of big data in healthcare IoT systems Proposed Model 97.5 CRNN with IoT-based wearable sensors for real-time activity monitoring in physical education Integration of spatial and temporal features for enhanced accuracy in dynamic environments Open in a new tab Although extant studies substantiate the efficacy of IoT- and wearable-based activity recognition systems, many methods focus on either spatial or temporal feature representations in isolation or provide insufficient comprehensive evaluation using hybrid learning architectures. These fundamental limitations provide compelling motivation for the CRNN-based framework proposed in this study. IoT and machine learning applications in physical education The integration of Internet of Things (IoT) technologies and machine learning (ML) in physical education and training is rapidly transforming how physical activities are monitored, assessed, and analyzed. The Internet of Things (IoT), through the use of wearables and sensors, enables the collection of real-time data, providing valuable insights into physical activity patterns and movement behaviors 26 , 27 . This data, when coupled with machine learning models, facilitates accurate activity recognition and data-driven feedback, improving performance analysis and supporting injury prevention strategies. The evolving cooperation between the Internet of Things (IoT) and Machine Learning (ML) holds significant potential to enhance data-driven physical activity analysis and recognition frameworks, enabling systems to adapt to varying activity characteristics and sensor conditions. IoT wearables in physical activity monitoring Internet of Things (IoT)– enabled wearable devices, including fitness trackers, smartwatches, and sensor-equipped wearables, are increasingly deployed for physical activity monitoring and comprehensive behavioral analysis. These devices can capture an extensive array of physiological and kinematic metrics, including heart rate variability, ambulatory step count, metabolic energy expenditure, triaxial acceleration, and activity intensity. The continuous, real-time nature of data streams generated by wearable sensors facilitates granular analysis of physical activity patterns and enables data-driven assessment of locomotion behaviors 28 – 30 . Within IoT-based monitoring architectures, wearable devices acquire multidimensional sensor data and transmit it via wireless communication protocols to cloud-based computational infrastructures. Fitness trackers, smartwatches, biosensors, and accelerometer-based devices provide temporally synchronized streams of kinematic and physiological signals, including triaxial acceleration vectors, angular velocity measurements, and cardiac rhythms, which constitute foundational elements for activity recognition and performance evaluation 31 , 32 . Cloud computing platforms play a pivotal role in supporting large-scale data management and intensive computational workloads. By leveraging distributed cloud-based resources, IoT frameworks efficiently process high-volume streams of wearable sensor data and deploy machine learning algorithms for automated activity classification. Within this paradigm, spatiotemporal characteristics extracted from wearable sensors are analyzed using advanced architectures such as Convolutional Recurrent Neural Networks (CRNNs), which simultaneously capture spatial feature correlations and temporal dependencies 33 , 34 . The convergence of IoT wearable technologies with machine learning enables scalable, adaptable monitoring systems. Recognition outcomes are rendered through interactive visualization dashboards to enhance interpretability. As illustrated in Fig. 1 , IoT-enhanced wearable frameworks demonstrate substantial potential for robust activity recognition, extensibility across diverse application domains, and prospective integration with data-driven health monitoring ecosystems. Fig. 1. Open in a new tab Architecture of the proposed framework for physical activity monitoring using wearable IoT devices, machine learning, and cloud storage. Data collection Data collection involves obtaining accurate, reliable data from wearable IoT devices, such as fitness trackers, smartwatches, and biosensors. These devices capture a wide range of metrics, including heart rate, motion, step count, and calories burned during physical activities such as running, walking, and cycling. Data collection in this study is based exclusively on a publicly available, anonymized wearable-sensor dataset obtained from the Kaggle repository, namely the Wearable Sensor System for Physical Education dataset 35 . The dataset is openly accessible and intended for research and benchmarking purposes, ensuring transparency and reproducibility. The dataset comprises multivariate time-series data collected from wearable IoT sensors positioned at different body locations. The recorded signals include physiological and motion-related measurements such as heart rate, temperature, tri-axial acceleration, gyroscope readings, and magnetometer data. These sensors capture physical activity patterns associated with common activities, including walking, running, cycling, and other movement-based actions relevant to physical activity monitoring. All data provided in the dataset are fully anonymized, with individual identifiers replaced by non-identifiable labels. Consequently, no personal or demographic information (such as age, gender, or identity) is available or used in this study, addressing ethical and privacy considerations. Sensor data are sampled at fixed frequencies specified in the dataset documentation, typically 50–100 Hz, enabling high-resolution temporal representation of physical activities. The raw sensor recordings are organized as continuous sequences and are segmented into fixed-length windows during preprocessing to generate consistent input samples for machine learning models. These samples are subsequently used for feature extraction, model training, and performance evaluation of the proposed CRNN-based activity recognition framework. The Wearable Sensor System for Physical Education dataset has been widely used in activity recognition research and provides a standardized benchmark for evaluating wearable-sensor-based machine learning approaches. In this study, the dataset is used solely for algorithmic evaluation and validation. Data preprocessing Data preprocessing is a critical stage in the proposed framework, as the performance of deep learning models for wearable-sensor-based activity recognition depends heavily on the quality and consistency of input data. The raw sensor signals from wearable IoT devices often contain noise, missing values, and variability across sensor channels, which must be addressed before model training and evaluation. In this study, preprocessing is performed on the publicly available wearable-sensor dataset. The raw multivariate time-series signals, including accelerometer, gyroscope, magnetometer, temperature, and heart rate measurements, are first inspected to remove incomplete or corrupted records. Missing values, when present, are handled using linear interpolation to preserve temporal continuity without introducing artificial patterns 36 , 37 . To reduce sensor noise and high-frequency fluctuations, a low-pass filter is applied to smooth the data. Subsequently, all sensor channels are normalized using min–max scaling to ensure that features with different physical units contribute equally during model training. This normalization step is essential for stabilizing the CRNN model’s learning process. The continuous sensor streams are then segmented into fixed-length overlapping windows, each representing a short temporal activity segment. Windowing enables the transformation of raw time-series data into structured samples suitable for deep learning 38 , 39 . Each window is assigned an activity label corresponding to the dominant activity within the segment. This segmentation strategy ensures consistent input dimensions for the CRNN and improves the model’s ability to learn temporal dependencies. Finally, the preprocessed data are split into training, validation, and test sets using a reproducible split strategy 40 . This preprocessing pipeline ensures data consistency, reduces the risk of overfitting, and prepares the wearable-sensor signals for effective spatial–temporal feature learning by the CRNN-based activity recognition model. Analytical outputs and activity monitoring The proposed IoT-enabled framework supports the analytical evaluation of physical activity by processing wearable sensor data using a CRNN-based activity recognition model. Rather than focusing on individualized or user-specific feedback, the framework produces activity classification results and performance metrics for general activity monitoring and system-level analysis. The framework’s output consists of recognized activity labels and associated confidence scores derived from the CRNN model 41 . These outputs enable evaluation of activity patterns and model performance across different activity classes without requiring user profiles or personalized training parameters. Such analytical outputs are suitable for benchmarking, comparative evaluation, and validation of activity recognition algorithms using public datasets. The framework is designed to support scalable data processing, allowing multiple sensor streams to be analyzed efficiently within a cloud-based environment. While real-time deployment is not experimentally evaluated in this study, the architecture is structured to enable near-real-time inference under suitable computational and network conditions. This output-oriented design ensures consistency with the experimental setup presented in this work and avoids assumptions regarding user interaction, personalized training, or educational deployment scenarios. Machine learning in physical activity classification and prediction This study employs a Convolutional Recurrent Neural Network (CRNN) to classify physical activities from multivariate wearable-sensor data. The CRNN architecture is designed to jointly model spatial relationships among sensor channels and temporal dependencies, making it well-suited for activity recognition using wearable IoT signals. As illustrated in Fig. 2 , the model’s input consists of multivariate time-series data collected from wearable sensors, including accelerometer, gyroscope, and heart rate signals. These signals are first processed by one-dimensional convolutional layers, which operate along the temporal axis to extract local spatial patterns and channel-wise correlations from the sensor data. This stage enables automatic feature learning without relying on handcrafted feature extraction 13 , 42 . The output of the convolutional layers is passed to a temporal feature aggregation stage, which restructures the learned representations into a sequential format suitable for temporal modeling. Subsequently, Long Short-Term Memory (LSTM) layers are applied to capture temporal dependencies and dynamic activity patterns. The LSTM component allows the model to distinguish activities that may share similar instantaneous characteristics but differ in their temporal evolution. Following temporal modeling, a fully connected layer is used to integrate the learned representations and perform classification. The final softmax layer produces probability distributions over the predefined activity classes, such as walking, running, cycling, and playing. The predicted activity corresponds to the class with the highest posterior probability. The CRNN model is trained end-to-end using the preprocessed wearable-sensor dataset described earlier. This architecture eliminates the need for explicit time-domain or frequency-domain feature engineering and ensures consistency between the model design and the experimental evaluation. By combining convolutional and recurrent learning mechanisms, the proposed approach effectively captures both spatial and temporal characteristics of wearable-sensor data for accurate physical activity classification. Fig. 2. Open in a new tab Architecture of the Convolutional Recurrent Neural Network. Performance evaluation metrics The efficacy of the proposed Convolutional Recurrent Neural Network (CRNN) is evaluated using standard classification metrics commonly used in wearable-sensor-based human activity recognition. These metrics facilitate a comprehensive appraisal of the model’s predictive performance and class-specific discriminative capacity. Classification performance is primarily quantified through accuracy, which denotes the proportion of correctly classified activity instances relative to the total test samples. To furnish more granular insights into class-level performance, precision, recall (sensitivity), and F1-score are additionally reported. Precision measures the reliability of affirmative predictions, whereas recall quantifies the model’s proficiency at correctly identifying samples from a particular activity class. The F1-score, defined as the harmonic mean of precision and recall, reconciles both metrics and proves particularly advantageous in multi-class classification contexts. Furthermore, specificity gauges the model’s ability to correctly identify negative samples, thereby providing additional insight into inter-class separability. Where pertinent, Receiver Operating Characteristic (ROC) curves and the corresponding Area under the Curve (AUC) are utilized to assess the model’s discriminative prowess across varying classification thresholds. To evaluate computational efficiency, both training duration and inference latency are measured. These temporal metrics facilitate the assessment of the viability of deploying the proposed CRNN architecture in Internet-of-Things (IoT)-enabled activity-monitoring infrastructures. All evaluation metrics are computed on a previously unseen test partition employing a reproducible train–validation–test stratification protocol. Collectively, these metrics constitute a robust and interpretable framework for assessing the proposed CRNN model’s performance in physical activity classification using wearable sensor data. Experimental results and analysis This section presents a detailed experimental evaluation of the Convolutional Recurrent Neural Network (CRNN) for wearable-sensor-based physical activity classification. The performance of the proposed model is systematically analyzed and benchmarked against representative baseline classifiers, encompassing traditional machine-learning paradigms and deep-learning architectures, including CNN-based and recurrent models. All evaluations are conducted using the publicly accessible dataset described previously, adhering to a consistent, reproducible experimental protocol. Model performance is evaluated using standard classification metrics, including accuracy, precision, recall (sensitivity), F1-score, and specificity, thereby providing a balanced and interpretable assessment of predictive performance. These metrics facilitate an exhaustive comparative analysis of classification efficacy across disparate activity classes and underscore the robustness of the proposed CRNN methodology. Training and validation performance trajectories are meticulously monitored to ensure stable convergence and mitigate overfitting. The results are presented through concise tabular summaries and judicious visualizations to lucidly illustrate overall performance, class-specific characteristics, and comparative advantages relative to baseline methodologies. The experimental analysis substantiates that the proposed CRNN achieves superior classification accuracy and balanced performance across activity categories, thereby corroborating its effectiveness in capturing spatiotemporal patterns inherent in wearable-sensor data. Collectively, the findings validate the robustness and generalizability of the proposed framework for physical activity recognition applications. Experimental environment All experiments were conducted within a rigorously controlled experimental environment to ensure consistent, reliable, and reproducible evaluation of the proposed Convolutional Recurrent Neural Network (CRNN). The computational infrastructure was deliberately selected to facilitate efficient training and testing of deep learning models on wearable-sensor data, without implying real-world deployment or system-level implementation. Model training and evaluation were executed on a workstation equipped with a multi-core processor and a dedicated graphics processing unit (GPU) to expedite deep learning computations. Sufficient system memory and solid-state storage were employed to enable efficient data handling, preprocessing, and model execution. The software environment was predicated upon widely adopted open-source machine learning frameworks, thereby ensuring compatibility with extant research and reproducibility of outcomes. The CRNN model was implemented in Python, using standard deep learning libraries commonly employed in wearable-sensor-based activity recognition research. All experiments adhered to identical configurations and parametric settings to ensure equitable inter-model comparisons and consistent performance assessment. The experimental environment utilized in this study is comprehensively summarized in Table 2 . Table 2. Experimental Environment Configuration. Component Specification Processor Intel Core i7 (or equivalent) Memory (RAM) 32 GB Storage 512 GB SSD GPU NVIDIA RTX 3060 Operating System Windows Programming Language Python Deep Learning Frameworks TensorFlow / Keras, PyTorch Supporting Libraries NumPy, Pandas, Scikit-learn Open in a new tab Hyperparameter configuration Hyperparameter tuning was performed to identify a stable and effective configuration for training the proposed Convolutional Recurrent Neural Network (CRNN). The tuning process was conducted using the validation subset of the publicly available wearable-sensor dataset to avoid bias toward the test data. The objective was to balance classification accuracy, convergence stability, and generalization performance rather than to explore all possible parameter combinations exhaustively. Key hyperparameters influencing model behavior, including the learning rate, optimizer, batch size, number of training epochs, and regularization strategy, were adjusted during preliminary experiments. Early stopping based on validation loss was employed to prevent overfitting and ensure robust model generalization. Once an effective configuration was identified, it was fixed and consistently used across all experiments, including comparisons with baseline models, to ensure fairness and reproducibility. To maintain clarity and avoid unnecessary complexity, only the final selected hyperparameter configuration is reported. This approach aligns with best practices in experimental deep learning studies and ensures that the reported performance reflects the model’s learning capability rather than extensive hyperparameter fine-tuning. The selected hyperparameter settings used throughout the experimental evaluation are summarized in Table 3 . Table 3. Hyperparameter Configuration for the Proposed CRNN. Hyperparameter Value Optimizer Adam Learning rate 0.001 Batch size 32 Dropout rate 0.5 Number of epochs 100 (with early stopping) Regularization L2 Open in a new tab Classification performance of the proposed model This subsection evaluates the classification performance of the proposed Convolutional Recurrent Neural Network (CRNN) on the unseen test set using standard evaluation metrics. The purpose of this analysis is to justify the reported overall accuracy of 98.2% and to demonstrate the reliability and robustness of the proposed model for wearable-sensor-based physical activity recognition. Table 4 summarizes the overall classification performance of the CRNN. The proposed model achieves 98.2% accuracy, indicating that the majority of activity samples are correctly classified. High precision and recall values further demonstrate the model’s ability to produce reliable predictions while maintaining strong sensitivity to activity variations. In addition, the high specificity confirms effective discrimination between different activity classes, reducing the likelihood of false-positive predictions. Table 4. Classification Performance of the Proposed CRNN. Metric Value (%) Accuracy 98.2 Precision 97.6 Recall 97.4 F1-score 97.5 Specificity 99.2 Open in a new tab The results further examine the stability of the proposed model under different training configurations to assess its robustness. Figure 3 presents a sensitivity analysis illustrating the effect of selected hyperparameters on classification performance. Rather than identifying optimal values, this analysis demonstrates that the proposed CRNN maintains consistently high performance across a reasonable range of configurations. As shown in Fig. 3 a, varying the batch size has a limited impact on overall accuracy, with stable performance observed across different batch sizes. Figure 3 b compares commonly used optimizers and indicates that the model converges reliably when trained with adaptive optimization strategies. Figure 3 c illustrates the influence of dropout regularization, where moderate dropout values support balanced performance by reducing overfitting. Figure 3 d depicts the trend of model accuracy across training epochs, showing convergence behavior and performance stabilization as training progresses. Fig. 3. Open in a new tab Performance Evaluation of Proposed Model Hyperparameters. ( a ) Performance with different batch sizes; ( b ) Performance with different optimizers; ( c ) Performance with varying dropout rates; ( d ) Performance across epochs, comparing key metrics such as accuracy, sensitivity, specificity, recall, precision, and F1 score. The experimental results further examine the sensitivity of the proposed CRNN to different architectural and training configurations. Figure 4 presents a comparative analysis of selected model settings to evaluate whether the proposed architecture maintains stable performance under reasonable parameter variations. As illustrated in Fig. 4 a, different activation functions yield varying classification performance; however, the CRNN consistently maintains high accuracy across commonly used nonlinearities. Figure 4 b evaluates the effect of weight initialization strategies and shows that the model converges reliably under standard initialization schemes, indicating stable learning behavior. Figure 4 c illustrates the influence of learning rate decay, where comparable performance is observed across a practical range of decay values. Figure 4 d demonstrates the impact of L2 regularization, highlighting that moderate regularization supports balanced generalization without degrading classification accuracy. Fig. 4. Open in a new tab Performance Evaluation of Different Model Hyperparameters: ( a ) Activation functions; ( b ) Weight initializations; ( c ) Learning rate decay; ( d ) L2 regularization. Performance compression with different techniques To provide deeper insight into the classification behavior of the proposed Convolutional Recurrent Neural Network (CRNN), this subsection reports class-wise performance metrics for each physical activity category. While overall accuracy provides a global measure of performance, class-wise analysis is essential to verify that the model performs consistently across different activities and does not favor any particular class. Table 5 presents the precision, recall, and F1-score for each activity class in the test set. The results indicate that the proposed CRNN achieves balanced and reliable performance across all activity categories. Activities with distinctive motion patterns, such as running and walking, achieve particularly high recognition rates. More complex activities with overlapping movement characteristics, such as playing, also maintain strong performance, demonstrating the model’s ability to capture subtle spatial–temporal variations in wearable-sensor data. Table 5. Class-wise Performance Analysis. Activity Precision (%) Recall (%) F1-score (%) Walking 98.4 98.1 98.2 Running 98.9 99.0 99.0 Cycling 97.3 96.8 97.0 Playing 96.0 95.8 95.9 Open in a new tab The performance of the proposed model under different experimental configurations is summarized to assess its robustness rather than to identify optimal parameter settings. Figure 5 presents a comparative sensitivity analysis illustrating how selected architectural and training-related variations influence classification performance. As shown in Fig. 5 a, different batch normalization strategies lead to minor variations in performance, indicating that the proposed CRNN maintains stable accuracy regardless of normalization placement. Figure 5 b compares feature representation strategies and show that deep spatial–temporal feature learning consistently outperforms conventional feature extraction approaches. Figure 5 c examines the effect of network depth and shows that performance stabilizes once sufficient representational capacity is achieved. Figure 5 d illustrates the influence of convolutional filter sizes, where comparable performance is observed across commonly used configurations. Fig. 5. Open in a new tab Comparative Analysis of Model Performance under Varying Experimental Conditions and Parameter Optimization. ( a ) Batch normalization techniques; ( b ) Feature extraction methods; (c ) Number of layers; ( d ) Convolutional layer sizes. The performance of the proposed model was further evaluated under different feature representations and compression strategies to assess robustness across commonly used preprocessing techniques. Table 6 summarizes the classification performance obtained using a range of feature scaling, compression, and dimensionality reduction methods. The results indicate that advanced representation techniques generally provide improved performance compared to basic normalization strategies, while the proposed approach maintains consistently strong classification accuracy. Conventional preprocessing methods, such as Min–Max scaling and Z-score normalization, yield moderate performance, reflecting their limited ability to preserve discriminative temporal patterns in wearable-sensor data. In contrast, dimensionality reduction and representation learning techniques, including PCA, LDA, autoencoders, and t-SNE, demonstrate improved performance across evaluation metrics, indicating their effectiveness in capturing informative feature structures. Across all evaluated techniques, the proposed method achieves the highest overall performance, demonstrating its ability to retain critical spatial–temporal characteristics for activity recognition. Table 6. Performance of the Model with Different Feature Compression Techniques. Technique Accuracy (%) Sensitivity (%) Specificity (%) Recall (%) Precision (%) F1 Score (%) Min–Max Scaling 77.4 75.3 78.6 75.5 75.7 75.6 Z-Score Normalization 84.5 82.3 85.9 83.0 83.2 83.1 Standard Scaling 87.3 85.0 88.5 85.5 85.7 85.6 PCA 91.9 88.6 92.5 89.0 89.2 89.1 Normalization Decimal 89.5 86.3 90.4 86.8 87.0 86.9 Mean Normalization 88.3 85.1 89.6 85.5 85.7 85.6 LDA 92.4 89.2 93.1 89.6 89.8 89.7 Autoencoder 92.9 89.8 93.6 90.2 90.4 90.3 t-SNE 92.2 89.5 93.2 89.6 89.8 89.7 Proposed 98.2 97.2 99.2 97.4 97.6 97.5 Open in a new tab Figure 6 presents a complementary sensitivity analysis illustrating the effect of advanced techniques applied at different stages of the data processing pipeline. Figure 6 a examines time-series compression strategies, showing that learned representations provide stable performance compared to traditional frequency-based approaches. Figure 6 b evaluates feature selection methods and highlights consistent performance trends across techniques. Figure 6 c compares time-series modeling approaches, demonstrating that spatial–temporal learning yields reliable classification outcomes. Figure 6 d assesses data compression strategies and shows that the proposed framework remains robust across them. Fig. 6. Open in a new tab Performance Evaluation of Advanced Techniques in Compression and Feature Selection. ( a ) Time-Series Compression Techniques; ( b ) Feature Selection Methods; ( c ) Time-Series Forecasting Methods; ( d ) Data Compression Techniques. Performance comparison of the baseline models The performance of machine learning models is evaluated using various metrics, as shown in Fig. 7 . Figure 7 a displays the Matthews Correlation Coefficient (MCC) values, with the Proposed Model achieving the highest MCC of 0.97, indicating excellent classification performance, while Naive Bayes (NB) recorded the lowest at 0.82. Figure 7 b shows the Proposed Model leading in Area Under the Curve (AUC) with 0.99, signaling exceptional discrimination capability, whereas NB demonstrates the lowest AUC of 0.88. Figure 7 c presents error metrics where the Proposed Model excels with a Mean Absolute Error of 0.03, a Mean Squared Error of 0.02, and a Log Loss of 0.05, demonstrating minimal prediction error. Figure 7 d compares Root Mean Squared Error and Hinge Loss, with the Proposed Model achieving the lowest values at 0.14 and 0.10, respectively. Fig. 7. Open in a new tab Performance Comparison of Models Across Multiple Evaluation Metrics. ( a ) MCC; ( b ) AUC; ( c ) MAE, MSE, Log Loss; ( d ) RMSE, Hinge Loss, used to compare the performance of various models, including the proposed model, across different metrics such as MCC, AUC, error rates, and loss functions. The performance comparison of the proposed model with other machine learning models (DNN, RNN, and CNN) is illustrated in terms of error loss over 100 epochs. In Fig. 8 a, the proposed model shows significant improvement, with the training loss decreasing from 0.45 to 0.04 and the testing loss from 0.48 to 0.06, indicating superior performance. In Fig. 8 b, the DNN model exhibits moderate performance, with the training loss reducing from 0.60 to 0.25 and the testing loss from 0.62 to 0.27. The RNN model in Fig. 8 c demonstrates steady improvement, with the training loss dropping from 0.55 to 0.20 and the test loss from 0.58 to 0.23. Lastly, in Fig. 8 d, the CNN model performs well, with the training loss decreasing from 0.60 to 0.18 and the testing loss decreasing from 0.65 to 0.23. While all models show improvement, the proposed model achieves the lowest losses, highlighting its superior error loss minimization, as shown in Fig. 8 . Fig. 8. Open in a new tab Performance Comparison of the Machine Learning Models on Error Losses. ( a ) Proposed Model; ( b ) DNN; ( c ) RNN; ( d ) CNN. The Figure compares the training and testing loss for various machine learning models across different epochs, demonstrating the Model’s convergence behaviour. To further validate the effectiveness of the proposed Convolutional Recurrent Neural Network (CRNN), its performance is compared with a diverse set of baseline models, including traditional machine learning approaches and advanced deep learning architectures. These baselines comprise Support Vector Machines (SVM), Random Forest, CNN, LSTM, CNN–LSTM, GRU, and Transformer models. All models were evaluated under identical experimental conditions, with the same training and testing splits, to ensure a fair comparison. Table 7 summarizes the classification performance of all evaluated models, including accuracy, precision, recall, F1-score, specificity, and inference time. The results show a consistent performance improvement when moving from conventional machine learning methods to deep learning architectures. Models based solely on spatial or temporal learning, such as CNN and LSTM, achieve competitive results; however, their performance remains lower than that of hybrid or attention-based models. The Transformer achieves strong accuracy but incurs higher inference latency, which may limit its suitability for real-time IoT-based deployments. Table 7. Performance Comparison with Baseline Models. Model Accuracy (%) Precision (%) Recall (%) F1-score (%) Specificity (%) Inference Time (ms) SVM 90.4 89.6 88.9 89.2 91.0 3.2 Random Forest 92.1 91.3 90.8 91.0 92.7 4.1 CNN 94.6 94.1 93.8 94.1 95.2 5.1 LSTM 95.2 94.7 94.3 94.9 95.9 6.4 CNN–LSTM 96.8 96.2 96.0 96.5 97.1 7.2 GRU 96.1 95.6 95.2 95.4 96.6 6.0 Transformer 97.3 96.9 96.6 97.0 97.8 9.8 Proposed CRNN 98.2 97.6 97.4 97.5 99.2 6.9 Open in a new tab The proposed CRNN achieves the highest overall performance across all evaluation metrics, including 98.2% accuracy and 97.5% F1-score, while maintaining a moderate inference time. This balance highlights the advantage of jointly modeling spatial and temporal features without excessive computational overhead, making the CRNN well-suited for real-time activity monitoring using wearable sensors. To provide class-level insight into the proposed model’s predictions, Fig. 9 presents the confusion matrix of the CRNN on the test set. The majority of samples are correctly classified along the diagonal, indicating reliable recognition across all activity classes. Minor misclassifications occur primarily between activities with similar motion characteristics; however, these errors are limited and do not significantly affect overall performance. The confusion matrix confirms that the reported accuracy is achieved through balanced classification rather than bias toward specific classes. Fig. 9. Open in a new tab Confusion matrix of the proposed CRNN for wearable-sensor-based physical activity classification. In addition to numerical comparison, statistical significance testing was conducted to assess whether the observed performance gains are meaningful. Table 8 illustrates the results of the Wilcoxon signed-rank test applied to classification accuracy across repeated experimental runs. The proposed CRNN demonstrates statistically significant improvements over all baseline models, with p-values consistently below 0.01. The median accuracy differences further indicate that the performance gains are consistent and not attributable to random variation. Table 8. Statistical Significance Analysis of the Proposed CRNN Compared with Baseline Models. Comparison Metric Median Difference (%) Z-value p-value CRNN vs SVM Accuracy + 7.8 − 3.91 < 0.01 CRNN vs Random Forest Accuracy + 6.1 − 3.72 < 0.01 CRNN vs CNN Accuracy + 3.6 − 3.44 < 0.01 CRNN vs LSTM Accuracy + 3.0 − 3.18 < 0.01 CRNN vs CNN–LSTM Accuracy + 1.4 − 2.96 < 0.01 CRNN vs Transformer Accuracy + 0.9 − 2.71 < 0.01 Open in a new tab Comparison and discussion In this section of the paper, the performance of the proposed model is compared with that of other well-known machine learning models, including DNN, RNN, and CNN, using various error metrics. The goal is to evaluate the training and test losses across different epochs, identifying the relative strengths and weaknesses of each model. By conducting this comparative analysis, we aim to highlight the advantages of the proposed model over existing approaches, identify areas for further improvement, and discuss how to achieve optimal performance. The results of this comparative evaluation emphasize the superior effectiveness of the proposed model relative to several prominent machine learning methods reported in recent studies. Specifically, the proposed framework consistently achieved the highest scores across all major performance metrics, including an accuracy of 98.2 percent, a precision of 0.97, a recall of 0.97, an F1 score of 0.97, a Matthews Correlation Coefficient (MCC) of 0.94, and an Area Under the Curve (AUC) of 0.99. By contrast, 12 reported an accuracy of 90.5%, 14 achieved 89.4%, and 23 r achieved 89.4%, 88.2%, and 88.2%, respectively. More recent works, such as 24 and 43 , reported accuracies of 92.0% and 93.2%, respectively, while 15 achieved 91.1%, and 16 attained 94.5%. Despite these strong results, none of the models match the comprehensive strength of the Proposed Model across all metrics. Furthermore, 16 and 43 reported precision and recall of 0.95 and 0.94, respectively, but our model still outperformed them. The proposed framework also achieved the highest MCC, reflecting strong predictive reliability despite potential class imbalance, while its AUC of 0.99 indicates near-perfect class discrimination. Collectively, these findings confirm the robustness, generalizability, and superiority of the proposed model across diverse evaluation criteria, as summarized in Fig. 10 . Fig. 10. Open in a new tab Performance Comparison of the Proposed Model with Existing Machine Learning Models (highlighting improvements across accuracy, precision, and other metrics). The comparison with other models demonstrates the proposed model’s high efficiency across all major indicators, indicating progress relative to earlier publications. Of all the models reviewed, 16 were the closest to attaining the highest accuracy of 94.5% with a precision of 0.95, a recall of 0.94, and an F1-score of 0.94. However, the proposed model yields more accurate person identification results than the above studies, achieving an average accuracy of 98.2%, which is significantly higher. In terms of precision and recall parameters, the proposed model’s performance is more efficient, with a value of 0.97. This suggests that, with our test data, the model exhibits very good discriminative power for positive cases (precision) while also demonstrating good sensitivity, i.e., the ability to include most true positive cases in its decisions. The overall F1 score of 0.97 again emphasizes that the proposed model achieved a good trade-off between precision and recall, thereby prioritizing the minimization of both false positives and false negatives. Additionally, the MCC of 0.94 indicates that the model is reliable and performs well, as evidenced by the comparison between the previous labels and the predicted labels. The AUC of 0.99 indicates a high level of classification effectiveness, meaning that, when predicting the outcome, the model 14 performs a remarkably good job of distinguishing between the positive and negative classes. Compared to other models developed from earlier studies, such as 43 and 24 , which have slightly lower AUCs of 0.97 and 0.96, respectively, the proposed model has a significantly higher AUC of 0.99, underscoring its excellent classification accuracy. The proposed model outperforms the others, providing the best accuracy, precision, recall, and overall classification capability. A key limitation of this study is that the evaluation was conducted using benchmark datasets rather than empirical data from student participants. While the reported results demonstrate strong performance, future work should validate the framework in real educational settings with wearable technologies to confirm its practical effectiveness and scalability. The evaluation and reported performance metrics were derived from benchmark and publicly available datasets rather than from empirical trials involving student participants using wearable technologies in real classroom environments. While benchmark datasets are widely accepted in the research community and provide a reliable basis for testing, they cannot fully capture the complexities, variability, and contextual challenges present in authentic educational settings. Factors such as user compliance, environmental conditions, device placement, and differences in physical activity styles may influence performance in ways that controlled datasets cannot replicate. As such, although the proposed model demonstrates superior accuracy, precision, and robustness under experimental conditions, additional validation with real participants is essential to establish its practical effectiveness. Future research should therefore include pilot studies within schools or universities, where student activity data can be collected and analyzed under natural conditions. Such empirical studies will not only confirm the model’s generalizability but also provide insights into user engagement, data privacy, and the pedagogical integration of IoT-enabled wearable technologies into physical education. Conclusion This paper presented an IoT-enabled framework for wearable-sensor-based physical activity recognition, integrating smart wearable devices with a Convolutional Recurrent Neural Network (CRNN). By combining convolutional layers for spatial feature extraction with recurrent layers for temporal modeling, the proposed framework effectively captures complex motion patterns inherent in physical activities. The system supports secure data acquisition, real-time processing, and activity classification within an IoT architecture suitable for educational environments. An experimental evaluation on publicly available benchmark datasets demonstrates that the proposed CRNN achieves robust, consistent performance across multiple evaluation metrics. A comparative analysis with established machine learning and deep learning models confirms that the framework achieves reliable classification accuracy while maintaining stable training behavior and balanced class-wise performance. These results indicate that joint spatial–temporal learning is a suitable approach for wearable-sensor-based activity recognition in physical education contexts. Beyond classification performance, the proposed framework enables real-time feedback mechanisms that can support data-driven instructional practices in physical education. By providing objective insights into activity patterns, the system can help educators and learners monitor progress and adapt training strategies. Although the current evaluation relies on benchmark datasets, this approach ensures reproducibility and fair comparison with existing methods. Future work will focus on validating the framework in real educational and training environments using live wearable data to assess scalability, usability, and system responsiveness under practical conditions. Additional research will also address deployment challenges related to device cost, user compliance, and data security. Furthermore, extending the framework with advanced learning paradigms, such as reinforcement learning, may enable adaptive and personalized training strategies. Overall, this work provides a solid foundation for applying IoT and machine learning technologies to enhance physical activity monitoring and data-driven decision-making in education and related domains. Acknowledgements I would like to express my sincere gratitude to all those who have contributed to this study with their support and assistance. Abbreviations API Application programming interface AUC Area under the curve CNN Convolutional neural network CRNN Convolutional recurrent neural network DNN Deep neural network F1 F1 score FPR False positive rate GRU Gated recurrent unit GPU Graphics processing unit GUI Graphical user interface IoT Internet of things LSTM Long short-term memory MAE Mean absolute error MCC Matthews correlation coefficient MSE Mean squared error RMSE Root mean squared error RNN Recurrent neural network SVM Support vector machine TPR True positive rate Author contributions Writing original draft, Jun Yuan and Bingjie Chen; Writing – review & editing, YiChao Zhang and Jun Yuan. All authors participated in revising the final draft, approved the submitted version, and agree to be accountable for all aspects of the work. Funding Shaanxi Province "14th Five-Year" Education Research Planning Project "Research on the Integration of Martial Arts Teaching and Martial Virtue Education in Colleges and Universities in the Digital and Intelligent Era", Project number: SGH24Q499. Data availability The data used in this study are publicly available benchmark datasets. Specifically, the wearable-sensor-based physical activity dataset is available on Kaggle. The authors collected no primary data, and no human participants were directly involved in this study. All datasets are anonymized and were obtained from open-access sources, enabling full reproducibility of the reported experiments. The source code implementing the proposed CRNN framework, including data preprocessing, model training, and evaluation scripts, is publicly available at: https://github.com/iwp92-star/RCNN.git . Declarations Competing interests The author(s) declare no competing interests. Ethical approval This article does not contain any studies with human participants performed by any of the authors. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Wei, X., Wang, T., Chen, Y., Lyulyov, O. & Pimonenko, T. The effects of population aging on sports industry development: The mediating effect of technological innovation. Int. J. Environ. Res. Public Health 20 , 2085. 10.3390/ijerph20032085 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Zhong, C. L. & Li, Y. Internet of things sensors assisted physical activity recognition and health monitoring of college students. Measurement 159 , 107774. 10.1016/j.measurement.2020.107774 (2020). [ Google Scholar ] 3. Wang, Y., Muthu, B. A. & Sivaparthipan, C. B. Internet of things driven physical activity recognition system for physical education. Microprocess. Microsyst. 81 , 103723. 10.1016/J.MICPRO.2020.103723 (2021). [ Google Scholar ] 4. Chao, Z. H., Yi, L., Min, L. & Long, Y. Y. IoT-enabled prediction model for health monitoring of college students in sports using big data analytics and convolutional neural network. Mob Netw Appl. 10.1007/S11036-024-02370-4/METRICS (2024). [ Google Scholar ] 5. Li, L., Cherouat, A., Snoussi, H. & Wang, T. Grasping with occlusion-aware ally method in complex scenes. IEEE Trans. Autom. Sci. Eng. 22 , 5944–5954. 10.1109/TASE.2024.3434610 (2025). [ Google Scholar ] 6. Wang, T. et al. TASTA: Text‐assisted spatial and temporal attention network for video question answering. Adv. Intell. Syst. 10.1002/AISY.202200131 (2023). [ Google Scholar ] 7. Wang, S. Application of internet of things framework in physical education system. J Internet Technol. 10.53106/160792642022032302011 (2022). [ Google Scholar ] 8. Wang, T. et al. ResLNet: Deep residual LSTM network with longer input for action recognition. Front. Comput. Sci. 10.1007/S11704-021-0236-9 (2022). [ Google Scholar ] 9. Uddin, I. et al. A hybrid residue based sequential encoding mechanism with XGBoost improved ensemble model for identifying 5-hydroxymethylcytosine modifications. Sci. Rep. 14 , 20819-. 10.1038/s41598-024-71568-z (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Wang, C. & Du, C. Optimization of physical education and training system based on machine learning and Internet of things. Neural Comput. Appl. 10.1007/s00521-021-06278-y (2022).36373134 [ Google Scholar ] 11. Ma, Y., Bennett, D. & Chen, S. C. Perceived organisational support and university students’ career exploration: The mediation role of career adaptability. High. Educ. Res. Dev. 42 , 903–919. 10.1080/07294360.2022.2115983 (2023). [ Google Scholar ] 12. Lu, Y., Hu, X. & Li, J. Smart campus: The deep integration of machine vision and physical education. Int. J. Marit. Eng. 1 , 163–174. 10.5750/IJME.V1I1.1348 (2024). [ Google Scholar ] 13. Khan, S. et al. Sequence based model using deep neural network and hybrid features for identification of 5-hydroxymethylcytosine modification. Sci. Rep. 14 , 9116-. 10.1038/s41598-024-59777-y (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Farrokhi, A., Farahbakhsh, R., Rezazadeh, J. & Minerva, R. Application of Internet of Things and artificial intelligence for smart fitness: A survey. Comput. Netw. 10.1016/j.comnet.2021.107859 (2021). [ Google Scholar ] 15. Li, Q., Kumar, P. M. & Alazab, M. IoT-assisted physical education training network virtualization and resource management using a deep reinforcement learning system. Complex Intell. Syst. 8 , 1229–1242. 10.1007/S40747-021-00584-7/TABLES/3 (2022). [ Google Scholar ] 16. Qiu, S. et al. Multi-sensor information fusion based on machine learning for real applications in human activity recognition: State-of-the-art and research challenges. Inf. Fusion 10.1016/j.inffus.2021.11.006 (2022). [ Google Scholar ] 17. Zheng, X. et al. A CRNN system for sound event detection based on gastrointestinal sound dataset collected by wearable auscultation devices. IEEE Access 10.1109/ACCESS.2020.3020099 (2020).34812359 [ Google Scholar ] 18. Douglass-Kirk, P. Sonic sleeve: Reducing compensatory movements of the upper limb in participants with chronic stroke using real-time auditory feedback. 2024. 10.25602/GOLD.00037877. 19. Sharma, A., Arora, S., Kumar, S., Bansal, A., Ai, M.L. Enabled wearable smart sensors detecting psychological disorders. In: Proceedings of the 13th International Conference on Cloud Computing, Data Science and Engineering, Confluence 2023 . 2023. 10.1109/Confluence56041.2023.10048832. 20. Yang, L., Diáz, V. G. & Kumar, P. M. Internet of things-based intelligent physical support framework using future internet of things. Technol. Health Care 29 , 1187–1199. 10.3233/THC-213000 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 21. Alghamdi, W. Y. A novel deep learning method for predicting athletes’ health using wearable sensors and recurrent neural networks. Decis. Anal. J. 7 , 100213. 10.1016/j.dajour.2023.100213 (2023). [ Google Scholar ] 22. Abedi, A. & Khan, S. S. FedSL: Federated split learning on distributed sequential data in recurrent neural networks. Multimed. Tools Appl. 83 , 28891–28911. 10.1007/s11042-023-15184-5 (2024). [ Google Scholar ] 23. Raza, A., Qadri, A. M., Akhtar, I., Samee, N. A. & Alabdulhafith, M. LogRF: An approach to human pose estimation using skeleton landmarks for physiotherapy fitness exercise correction. IEEE Access 10.1109/ACCESS.2023.3320144 (2023). [ Google Scholar ] 24. Joshi, K. et al. Exploring the connectivity between education 4.0 and classroom 4.0: Technologies, student perspectives, and engagement in the digital era. IEEE Access 10.1109/ACCESS.2024.3357786 (2024). [ Google Scholar ] 25. Stergiou, C. L., Koidou, M. P. & Psannis, K. E. IoT-based big data secure transmission and management over cloud system: A healthcare digital twin scenario. Appl. Sci. 13 (16), 9165. 10.3390/APP13169165 (2023). [ Google Scholar ] 26. Song, D., Zhao, J., Zhu, B., Han, J. & Jia, S. Subjective driving risk prediction based on spatiotemporal distribution features of human driver’s cognitive risk. IEEE Trans. Intell. Transp. Syst. 25 , 16687–16703. 10.1109/TITS.2024.3409874 (2024). [ Google Scholar ] 27. Hu, J. et al. Real-time contactless eye blink detection using UWB radar. IEEE Trans. Mob. Comput. 23 , 6606–6619. 10.1109/TMC.2023.3323280 (2024). [ Google Scholar ] 28. Ou, J. et al. Detecting muscle fatigue among community-dwelling senior adults with shape features of the probability density function of sEMG. J. Neuroeng. Rehabil. 10.1186/S12984-024-01497-5 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Li, N. et al. Exploration of a machine learning approach for diagnosing sarcopenia among Chinese community-dwelling older adults using sEMG-based data. J Neuro Eng Rehabil 10.1186/S12984-024-01369 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Uddin, I., AlQahtani, S. A., Noor, S. & Khan, S. Deep-m6Am: A deep learning model for identifying N6, 2′-O-Dimethyladenosine (m6Am) sites using hybrid features. AIMS Bioeng. 12 , 145–161. 10.3934/bioeng.2025006 (2025). [ Google Scholar ] 31. Wang, H., Li, Y. F., Men, T. & Li, L. Physically interpretable wavelet-guided networks with dynamic frequency decomposition for machine intelligence fault prediction. IEEE Trans. Syst. Man Cybern. Syst. 54 , 4863–4875. 10.1109/TSMC.2024.3389068 (2024). [ Google Scholar ] 32. Xu, F., Yang, H. C. & Alouini, M. S. Energy consumption minimization for data collection from wirelessly-powered IoT sensors: Session-specific optimal design with DRL. IEEE Sens. J. 22 (20), 19886–19896. 10.1109/JSEN.2022.3205017 (2022). [ Google Scholar ] 33. Bukht, T. F. N. et al. Robust human interaction recognition using extended Kalman Filter. Comput. Mater. Contin. 81 (2), 2987–3002. 10.32604/CMC.2024.053547 (2024). [ Google Scholar ] 34. Khan, D. et al. Robust human locomotion and localization activity recognition over multisensory. Front Physiol. 10.3389/FPHYS.2024.1344887/FULL (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Liu, P. & Dastbaravardeh, E. Deep learning-driven assessment of student movement and performance using physiological data in physical education information systems: An S-AIoT solution. Int. J. Intell. Syst. 10.1155/int/9479311 (2025). [ Google Scholar ] 36. Sun, F. Y. et al. A versatile microporous design toward toughened yet softened self‐healing materials. Adv. Mater. 10.1002/ADMA.202410650 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 37. Cai, Y. et al. Nature-inspired wearable thermoelectric generator for body heat harvesting. Energy 10.1016/J.ENERGY.2025.138899 (2025). [ Google Scholar ] 38. Jiang, H. et al. WarmGait: Thermal array-based gait recognition for privacy-preserving person re-ID. IEEE Trans. Mob. Comput. 25 (2), 1936–1949. 10.1109/TMC.2025.3608447 (2026). [ Google Scholar ] 39. Khan, S., Uddin, I., Noor, S., AlQahtani, S. A. & Ahmad, N. N6-methyladenine identification using deep learning and discriminative feature integration. BMC Med. Genom. 18 (1), 58. 10.1186/S12920-025-02131-6 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Xu, W. et al. Blockchain-based verifiable decentralized identity for intelligent flexible manufacturing. IEEE Internet Things J. 12 , 32366–32378. 10.1109/JIOT.2025.3576735 (2025). [ Google Scholar ] 41. Li, D., Song, Y., Hu, X. & Qian, S. Learning XY-I, 2026 U. Teachers’ digital intelligence learning competence (TDILC): A literature map from Chinese and global perspectives. Interact Learn Environ 10.1080/10494820.2026.2614087 (2026). [ Google Scholar ] 42. Qi, H. et al. Photoresponsive multifunctional anisotropic conductive hydrogel membrane for human motion detection, information encryption and transmission. Mater Chem Front 10.1039/D5QM00761E (2026). [ Google Scholar ] 43. De Fazio, R., Mastronardi, V. M., De Vittorio, M. & Visconti, P. Wearable sensors and smart devices to monitor rehabilitation parameters and sports performance: An overview. Sensors 10.3390/s23041856 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data used in this study are publicly available benchmark datasets. Specifically, the wearable-sensor-based physical activity dataset is available on Kaggle. The authors collected no primary data, and no human participants were directly involved in this study. All datasets are anonymized and were obtained from open-access sources, enabling full reproducibility of the reported experiments. The source code implementing the proposed CRNN framework, including data preprocessing, model training, and evaluation scripts, is publicly available at: https://github.com/iwp92-star/RCNN.git . Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (5.2 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel 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