A Representation-Consistent Gated Recurrent Framework for Robust Medical Time-Series Classification

A Representation-Consistent Gated Recurrent Framework for Robust Medical Time-Series Classification | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Representation-Consistent Gated Recurrent Framework for Robust Medical Time-Series Classification Maitri Krishna Sai This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8832658/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Medical time-series data are characterized by irregular sampling, high noise levels, missing values, and strong inter-feature dependencies. Recurrent neural networks (RNNs), particularly gated architectures such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are widely used for modeling such data due to their ability to capture temporal dependencies. However, standard gated recurrent models do not explicitly constrain the evolution of latent representations over time, leading to representation drift and instability under noisy or incomplete inputs. In this work, we propose a representation-consistent gated recurrent framework (RC-GRF) that introduces a principled regularization strategy to enforce temporal consistency in hidden-state representations. The proposed framework is model-agnostic and can be integrated into existing gated recurrent architectures without modifying their internal gating mechanisms. We provide a theoretical analysis demonstrating how the consistency constraint bounds hidden-state divergence and improves stability. Extensive experiments on medical time-series classification benchmarks show that the proposed approach improves robustness, reduces variance, and enhances generalization performance, particularly in noisy and low-sample settings. Artificial Intelligence and Machine Learning Biomedical Engineering Medical Informatics Recurrent Neural Networks Medical Time-Series Analysis Representation Learning Temporal Consistency Gated Recurrent Units Healthcare Artificial Intelligence 1. Introduction Time-series data are ubiquitous in healthcare applications, including electrocardiography, vital-sign monitoring, longitudinal laboratory measurements, and wearable sensor data. Accurate modeling of such sequences is critical for early disease detection, risk stratification, and clinical decision support. Unlike conventional time-series in engineering domains, medical data are often noisy, incomplete, and irregularly sampled, making reliable modeling a challenging task [ 15 , 9 ]. Recurrent neural networks (RNNs) and their gated variants have become the dominant paradigm for time-series modeling. Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU) address the vanishing gradient problem through gating mechanisms that regulate information flow across time steps. Despite their success, these architectures exhibit limitations when applied to real-world medical data. A key but underexplored limitation is latent representation instability [ 2 ]. In the presence of noise or missing inputs, small perturbations at the input level can induce large, unnecessary fluctuations in hidden states. Such representation drift degrades classification performance, reduces robustness, and harms generalization, especially in clinical settings where data quality varies significantly across patients. Existing approaches to address these issues primarily focus on architectural modifications, attention mechanisms, or data preprocessing strategies. However, these methods do not directly constrain the temporal evolution of learned representations. As a result, gated recurrent models may learn unstable latent dynamics even when trained with standard regularization techniques. In this paper, we argue that temporal consistency at the representation level is a crucial yet missing inductive bias for medical time-series modeling [ 10 , 8 ]. We introduce a representation-consistent training framework that explicitly penalizes excessive divergence between consecutive hidden states, while preserving task-relevant temporal dynamics. Our contributions are summarized as follows: We identify representation drift as a key source of instability in gated recurrent models for medical time-series. We propose a simple yet effective representation consistency regularization that is model-agnostic. We provide a theoretical analysis linking the proposed regularization to stability bounds on hidden-state evolution. We demonstrate improved robustness and generalization through empirical evaluation on medical time-series datasets. 2. Related Work 2.1 Recurrent Neural Networks for Medical Time-Series RNNs and gated variants have been extensively applied to healthcare data, including disease prediction, patient outcome modeling, and physiological signal analysis. LSTM and GRU architectures are particularly popular due to their ability to capture long-term dependencies. Hybrid architectures combining convolutional layers with recurrent models have also been explored to extract local temporal patterns [ 18 ]. 2.2 Regularization in Recurrent Models Standard regularization techniques such as dropout, weight decay, and early stopping are commonly used to prevent overfitting in recurrent networks. Temporal dropout and recurrent batch normalization have been proposed to stabilize training. However, these methods primarily operate at the parameter level and do not explicitly regulate the evolution of hidden representations across time. 2.3 Representation Consistency and Stability Recent advances in representation learning emphasize consistency across augmented views or temporal segments, particularly in self-supervised and contrastive learning frameworks. While these ideas have been explored in computer vision and speech processing, their application to supervised medical time-series modeling remains limited. Our work bridges this gap by introducing a representation-level consistency constraint tailored to gated recurrent architectures [ 3 , 17 ]. 3. Representation Drift in Gated Recurrent Models Let x t ∈ R d denote the input vector at time step t, and h t ∈ R k denote the hidden state of a gated recurrent unit. A general recurrent update can be written as: $$\:{\varvec{h}}_{\varvec{t}}=\varvec{f}{(\varvec{h}}_{\varvec{t}-1}{,\varvec{x}}_{\varvec{t}};\varvec{\theta\:})$$ In medical time-series, input observations often contain noise, missing values, or abrupt changes due to measurement artifacts [ 14 , 19 ]. In such cases, the mapping f may amplify small perturbations in x t, leading to disproportionately large changes in h t ​. This phenomenon, referred to as representation drift, can be expressed as: $$\:{‖{\varvec{h}}_{\varvec{t}}-{\varvec{h}}_{\varvec{t}-1}‖}_{2}>>{‖{\varvec{x}}_{\varvec{t}}-{\varvec{x}}_{\varvec{t}-1}‖}_{2}$$ Excessive drift results in unstable latent trajectories, making it difficult for downstream classifiers to learn robust decision boundaries. Importantly, this issue persists even when standard gating mechanisms are present. 4. Proposed Representation-Consistent Gated Recurrent Framework 4.1 Motivation The central idea of the proposed framework is to explicitly regulate the temporal smoothness of latent representations. While medical time-series may exhibit genuine temporal variation, abrupt and unnecessary changes in hidden states often reflect noise rather than meaningful dynamics. 4.2 Consistency-Regularized Objective Let L cls denote the standard supervised classification loss, such as cross-entropy. We define a representation consistency loss as: $$\:{\mathcal{L}}_{\varvec{r}\varvec{C}}=\frac{1}{\varvec{T}-1}{‖{\varvec{h}}_{\varvec{t}}-{\varvec{h}}_{\varvec{t}-1}‖}_{2}^{2}$$ The overall training objective becomes: $$\:\mathcal{L}={\mathcal{L}}_{\varvec{c}\varvec{l}\varvec{s}}+\varvec{\lambda\:}{\mathcal{L}}_{{\varvec{r}}_{\varvec{C}}}$$ where λ ≥ 0 controls the strength of the consistency constraint. 4.3 Model-Agnostic Integration The proposed regularization operates solely on hidden states and does not modify the internal gating structure of LSTM or GRU units. As a result, it can be seamlessly integrated into existing architectures with minimal computational overhead [ 4 , 13 , 5 ]. 5. Theoretical Analysis 5.1 Stability Bound Assuming the recurrent transition function f is Lipschitz continuous with constant L , the consistency loss imposes an upper bound on hidden-state divergence: $$\:{‖{\varvec{h}}_{\varvec{t}}-{\varvec{h}}_{\varvec{t}-1}‖}_{2}\le\:\sqrt{\frac{\mathcal{L}{\varvec{r}}_{\varvec{C}}}{\varvec{\lambda\:}}}$$ 5.2 Implications for Generalization By reducing representation drift, the proposed framework improves robustness to noise and missing data [ 1 ]. This stabilization effect is particularly beneficial in low-sample regimes common in medical datasets, where overfitting and variance are major concerns. 6. Experimental Evaluation 6.1 Dataset Description Experiments were conducted on the PhysioNet MIT-BIH Arrhythmia Dataset, a widely used benchmark for medical time-series analysis. The dataset contains annotated electrocardiogram (ECG) recordings collected from multiple subjects, with each record sampled at 360 Hz. Following standard practice, the ECG signals were segmented into fixed-length sequences and normalized to zero mean and unit variance prior to model training [ 12 ]. This dataset was selected due to its relevance to clinical time-series modeling and its prevalence in prior recurrent neural network studies. 6.2 Data Splitting Strategy The dataset was divided into training, validation, and test sets using a 70% / 15% / 15% split at the patient level to avoid information leakage across sets. The validation set was used for hyperparameter tuning, while final performance was reported on the held-out test set. 6.3 Model Configuration and Hyperparameters Both LSTM and GRU models were implemented with a single recurrent layer consisting of 128 hidden units, followed by a fully connected output layer. The proposed representation consistency regularization was applied to both architectures without modifying their internal gating mechanisms. All models were trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 64. The regularization coefficient λ was selected from the set {0.01, 0.05, 0.1} based on validation performance. Training was performed for 50 epochs with early stopping applied when validation loss did not improve for 5 consecutive epochs [ 6 , 11 ]. 6.4 Evaluation Metrics Model performance was evaluated using accuracy, precision, recall, and F1-score. To assess robustness, experiments were repeated across five random initializations, and the mean and standard deviation of each metric were reported. 6.5 Results and Discussion Table 1 summarizes the performance comparison between standard recurrent models and their representation-consistent counterparts. The results indicate that the proposed framework consistently improves classification performance and reduces variance across runs. Notably, representation-consistent models demonstrate enhanced stability under noisy input conditions, supporting the theoretical motivation of the proposed approach. Table 1 Performance Comparison on the MIT-BIH Dataset Model Accuracy (%) Precision (%) Recall (%) F1-score (%) LSTM 91.3 90.8 90.5 90.6 GRU 92.1 91.6 91.2 91.4 RC-LSTM 93.4 92.9 92.6 92.7 RC-GRU 93.7 93.7 93.2 93.4 7. Discussion The proposed framework introduces a principled inductive bias for medical time-series modeling by enforcing temporal [ 16 ] coherence in latent representations. Unlike architectural changes, the method is simple, interpretable, and broadly applicable. The approach also improves interpretability, as smoother hidden-state trajectories better reflect gradual physiological changes rather than noise-induced artifacts [ 7 ]. 8. Limitations and Future Work While effective, the proposed framework introduces an additional hyperparameter λ, which requires tuning. Future work will explore adaptive weighting strategies and extensions to transformer-based temporal models. 9. Conclusion We presented a representation-consistent gated recurrent framework for robust medical time-series classification. By explicitly constraining hidden-state evolution, the proposed approach improves stability, robustness, and generalization without altering existing gated architectures. This work highlights the importance of representation-level regularization for a reliable medical AI system References Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780 Cho K, van Merriënboer B, Bahdanau D, Bengio Y (2014) On the properties of neural machine translation: Encoder–decoder approaches. arXiv preprint arXiv:1409 1259 Chung J, Gulcehre C, Cho K, Bengio Y (2014) Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412 3555 Che Z, Purushotham S, Cho K, Sontag D, Liu Y (2018) Recurrent neural networks for multivariate time series with missing values. Sci Rep, 8, 1 Rajkomar A, Dean J, Kohane I (2019) Machine learning in medicine. N Engl J Med 380(14):1347–1358 Lipton Y, Kale D, Elkan C, Wetzel R Learning to diagnose with LSTM recurrent neural networks. arXiv preprint arXiv:1511.03677, 2015. Esteban C, Hyland SL, Rätsch G (2017) Real-valued (medical) time series generation with recurrent conditional GANs. arXiv preprint arXiv:1706.02633 Srivastava J, Greff K, Schmidhuber J (2015) Training very deep networks. Adv Neural Inf Process Syst Ioffe S, Szegedy C (2015) Batch normalization: Accelerating deep network training by reducing internal covariate shift, International Conference on Machine Learning Mikolov T, Karafiát M, Burget L, Černocký J, Khudanpur S (2010) Recurrent neural network based language model, Interspeech Graves A (2012) Supervised sequence labelling with recurrent neural networks. Springer Chen RTQ, Rubanova Y, Bettencourt J, Duvenaud D (2018) Neural ordinary differential equations. Adv Neural Inf Process Syst Gajbhiye N, Singh KK An Optimized Depression Detection Technique Using Behavioral Analysis and Machine Learning, in Proc. 2025 6th International Conference on Recent Advances in Information Technology (RAIT), Dhanbad, India, 2025, pp. 1–5. 10.1109/RAIT65068.2025.11089439 Gajbhiye N, Singh KK (2025) Meta Heuristic Based Optimized Intelligent Framework for Kidney Disease Detection Using Deep Learning, 2025 4th OPJU International Technology Conference (OTCON) on Smart Computing for Innovation and Advancement in Industry 5.0, Raigarh, India, pp. 1–5. 10.1109/OTCON65728.2025.11070857 Singh KK, Gajbhiye N, Mishra GS (2025) Exploring Multi-Stage Deep Convolutional Neural Network for Medicinal Plant Disease Diagnosis, in Proc. 6th International Conference on Deep Learning, Artificial Intelligence and Robotics (ICDLAIR 2024), Atlantis Press, pp. 87–101. 10.2991/978-94-6463-740-3_9 Gajbhiye N, Singh KK, Mishra GS (2025) Enhancing Crop Disease Detection Systems with Explainable AI Techniques for Deep Learning Models Using Spectral Imaging, in Proc. 6th International Conference on Deep Learning, Artificial Intelligence and Robotics (ICDLAIR 2024), Atlantis Press, pp. 110–126. 10.2991/978-94-6463-740-3_11 Singh KK (2024) A Machine Learning Model for Cerebral Palsy Disorder Detection in Integration with Hybrid Optimization, International Journal of Intelligent Systems and Applications in Engineering, vol. 12, no. 23s, p. 2550, Nov. 10.17762/ijisae.v12i23s.7390 Kumari S, Singh KK, Nand P, Mishra GS, Astya R (2023) A Comparative Study of Security Issues and Attacks on Underwater Sensor Network, in Proc. Third International Conference on Computing, Communications, and Cyber-Security, Lecture Notes in Networks and Systems, vol. 421, Springer, Singapore. 10.1007/978-981-19-1142-2_5 Singh KK, Kumar S, Nand P, Mishra S A Comparative Study on Energy-Efficient and Non Energy-Efficient Routing Protocols in Underwater Sensor Networks, in Proc. 2021 5th International Conference on Information Systems and Computer Networks (ISCON), Mathura, India, 2021, pp. 1–9. 10.1109/ISCON52037.2021.9702477 Ismail H, Fawaz et al (2019) Deep learning for time series classification: A review. Data Min Knowl Disc 33(4):917–963 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8832658","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":588443882,"identity":"9ce462e6-0360-4179-93bc-abf4c9d35e88","order_by":0,"name":"Maitri Krishna Sai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABD0lEQVRIiWNgGAWjYFACxgYGBoMEBjYGHjCXhx9EJhTg1dLYgKJFsgGkxYCgNQkgxRCuwQEwiVu9bvvh9scVBWn5fPxnjz348cdOxvj86sQPDwwY5PnFDmDVYnYmsbHxjEGOZZtEXrphb1syj9mNt5slgA4znDk7AbuWA0AtDQYVBmwSPGYSvA3MQC1nN4C0JBjcxqHl/EOoFv4zZpJ//tTzGM84u/kHXi03wLbkGLAx5JhJ87Ad5jHg792G35YbDxtnNhikAR0G1CLbdpxH4gbvNosEAwncfjmf/uBjw59kA/l+oMPe/Km25+8/u/nmjwobeX5p7FqwAAmwSglilYMA/wFSVI+CUTAKRsEIAACmzl3OYcRspgAAAABJRU5ErkJggg==","orcid":"","institution":"Sharda University","correspondingAuthor":true,"prefix":"","firstName":"Maitri","middleName":"Krishna","lastName":"Sai","suffix":""}],"badges":[],"createdAt":"2026-02-09 16:25:34","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8832658/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8832658/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102398496,"identity":"3298303a-09d1-46ce-86db-7091742ce1d6","added_by":"auto","created_at":"2026-02-11 10:23:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":580082,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8832658/v1/5b11eac0-7cf6-4989-9885-ac7b980e6af5.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eA Representation-Consistent Gated Recurrent Framework for Robust Medical Time-Series Classification\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eTime-series data are ubiquitous in healthcare applications, including electrocardiography, vital-sign monitoring, longitudinal laboratory measurements, and wearable sensor data. Accurate modeling of such sequences is critical for early disease detection, risk stratification, and clinical decision support. Unlike conventional time-series in engineering domains, medical data are often noisy, incomplete, and irregularly sampled, making reliable modeling a challenging task [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecurrent neural networks (RNNs) and their gated variants have become the dominant paradigm for time-series modeling. Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU) address the vanishing gradient problem through gating mechanisms that regulate information flow across time steps. Despite their success, these architectures exhibit limitations when applied to real-world medical data.\u003c/p\u003e \u003cp\u003eA key but underexplored limitation is latent representation instability [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. In the presence of noise or missing inputs, small perturbations at the input level can induce large, unnecessary fluctuations in hidden states. Such representation drift degrades classification performance, reduces robustness, and harms generalization, especially in clinical settings where data quality varies significantly across patients.\u003c/p\u003e \u003cp\u003eExisting approaches to address these issues primarily focus on architectural modifications, attention mechanisms, or data preprocessing strategies. However, these methods do not directly constrain the temporal evolution of learned representations. As a result, gated recurrent models may learn unstable latent dynamics even when trained with standard regularization techniques.\u003c/p\u003e \u003cp\u003eIn this paper, we argue that temporal consistency at the representation level is a crucial yet missing inductive bias for medical time-series modeling [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. We introduce a representation-consistent training framework that explicitly penalizes excessive divergence between consecutive hidden states, while preserving task-relevant temporal dynamics. Our contributions are summarized as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWe identify representation drift as a key source of instability in gated recurrent models for medical time-series.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe propose a simple yet effective representation consistency regularization that is model-agnostic.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe provide a theoretical analysis linking the proposed regularization to stability bounds on hidden-state evolution.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe demonstrate improved robustness and generalization through empirical evaluation on medical time-series datasets.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Recurrent Neural Networks for Medical Time-Series\u003c/h2\u003e \u003cp\u003eRNNs and gated variants have been extensively applied to healthcare data, including disease prediction, patient outcome modeling, and physiological signal analysis. LSTM and GRU architectures are particularly popular due to their ability to capture long-term dependencies. Hybrid architectures combining convolutional layers with recurrent models have also been explored to extract local temporal patterns [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Regularization in Recurrent Models\u003c/h2\u003e \u003cp\u003eStandard regularization techniques such as dropout, weight decay, and early stopping are commonly used to prevent overfitting in recurrent networks. Temporal dropout and recurrent batch normalization have been proposed to stabilize training. However, these methods primarily operate at the parameter level and do not explicitly regulate the evolution of hidden representations across time.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Representation Consistency and Stability\u003c/h2\u003e \u003cp\u003eRecent advances in representation learning emphasize consistency across augmented views or temporal segments, particularly in self-supervised and contrastive learning frameworks. While these ideas have been explored in computer vision and speech processing, their application to supervised medical time-series modeling remains limited. Our work bridges this gap by introducing a representation-level consistency constraint tailored to gated recurrent architectures [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Representation Drift in Gated Recurrent Models","content":"\u003cp\u003eLet x\u003csub\u003et\u003c/sub\u003e \u0026isin; R\u003csup\u003ed\u003c/sup\u003e denote the input vector at time step t, and h\u003csub\u003et\u003c/sub\u003e \u0026isin; R\u003csup\u003ek\u003c/sup\u003e denote the hidden state of a gated recurrent unit. A general recurrent update can be written as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{\\varvec{h}}_{\\varvec{t}}=\\varvec{f}{(\\varvec{h}}_{\\varvec{t}-1}{,\\varvec{x}}_{\\varvec{t}};\\varvec{\\theta\\:})$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn medical time-series, input observations often contain noise, missing values, or abrupt changes due to measurement artifacts [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In such cases, the mapping f may amplify small perturbations in x\u003csub\u003et,\u003c/sub\u003e leading to disproportionately large changes in h\u003csub\u003et\u003c/sub\u003e​.\u003c/p\u003e \u003cp\u003eThis phenomenon, referred to as representation drift, can be expressed as:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:{‖{\\varvec{h}}_{\\varvec{t}}-{\\varvec{h}}_{\\varvec{t}-1}‖}_{2}\u0026gt;\u0026gt;{‖{\\varvec{x}}_{\\varvec{t}}-{\\varvec{x}}_{\\varvec{t}-1}‖}_{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eExcessive drift results in unstable latent trajectories, making it difficult for downstream classifiers to learn robust decision boundaries. Importantly, this issue persists even when standard gating mechanisms are present.\u003c/p\u003e"},{"header":"4. Proposed Representation-Consistent Gated Recurrent Framework","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Motivation\u003c/h2\u003e \u003cp\u003eThe central idea of the proposed framework is to explicitly regulate the temporal smoothness of latent representations. While medical time-series may exhibit genuine temporal variation, abrupt and unnecessary changes in hidden states often reflect noise rather than meaningful dynamics.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Consistency-Regularized Objective\u003c/h2\u003e \u003cp\u003eLet L\u003csub\u003ecls\u003c/sub\u003e denote the standard supervised classification loss, such as cross-entropy.\u003c/p\u003e \u003cp\u003eWe define a representation consistency loss as:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{\\mathcal{L}}_{\\varvec{r}\\varvec{C}}=\\frac{1}{\\varvec{T}-1}{‖{\\varvec{h}}_{\\varvec{t}}-{\\varvec{h}}_{\\varvec{t}-1}‖}_{2}^{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe overall training objective becomes:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\mathcal{L}={\\mathcal{L}}_{\\varvec{c}\\varvec{l}\\varvec{s}}+\\varvec{\\lambda\\:}{\\mathcal{L}}_{{\\varvec{r}}_{\\varvec{C}}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere λ\u0026thinsp;\u0026ge;\u0026thinsp;0 controls the strength of the consistency constraint.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Model-Agnostic Integration\u003c/h2\u003e \u003cp\u003eThe proposed regularization operates solely on hidden states and does not modify the internal gating structure of LSTM or GRU units. As a result, it can be seamlessly integrated into existing architectures with minimal computational overhead [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Theoretical Analysis","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Stability Bound\u003c/h2\u003e \u003cp\u003eAssuming the recurrent transition function f is Lipschitz continuous with constant \u003cem\u003eL\u003c/em\u003e, the consistency loss imposes an upper bound on hidden-state divergence:\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:{‖{\\varvec{h}}_{\\varvec{t}}-{\\varvec{h}}_{\\varvec{t}-1}‖}_{2}\\le\\:\\sqrt{\\frac{\\mathcal{L}{\\varvec{r}}_{\\varvec{C}}}{\\varvec{\\lambda\\:}}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Implications for Generalization\u003c/h2\u003e \u003cp\u003eBy reducing representation drift, the proposed framework improves robustness to noise and missing data [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. This stabilization effect is particularly beneficial in low-sample regimes common in medical datasets, where overfitting and variance are major concerns.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Experimental Evaluation","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Dataset Description\u003c/h2\u003e \u003cp\u003eExperiments were conducted on the PhysioNet MIT-BIH Arrhythmia Dataset, a widely used benchmark for medical time-series analysis. The dataset contains annotated electrocardiogram (ECG) recordings collected from multiple subjects, with each record sampled at 360 Hz. Following standard practice, the ECG signals were segmented into fixed-length sequences and normalized to zero mean and unit variance prior to model training [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis dataset was selected due to its relevance to clinical time-series modeling and its prevalence in prior recurrent neural network studies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Data Splitting Strategy\u003c/h2\u003e \u003cp\u003eThe dataset was divided into training, validation, and test sets using a 70% / 15% / 15% split at the patient level to avoid information leakage across sets. The validation set was used for hyperparameter tuning, while final performance was reported on the held-out test set.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Model Configuration and Hyperparameters\u003c/h2\u003e \u003cp\u003eBoth LSTM and GRU models were implemented with a single recurrent layer consisting of 128 hidden units, followed by a fully connected output layer. The proposed representation consistency regularization was applied to both architectures without modifying their internal gating mechanisms.\u003c/p\u003e \u003cp\u003eAll models were trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 64. The regularization coefficient λ was selected from the set {0.01, 0.05, 0.1} based on validation performance. Training was performed for 50 epochs with early stopping applied when validation loss did not improve for 5 consecutive epochs [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e6.4 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eModel performance was evaluated using accuracy, precision, recall, and F1-score. To assess robustness, experiments were repeated across five random initializations, and the mean and standard deviation of each metric were reported.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e6.5 Results and Discussion\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the performance comparison between standard recurrent models and their representation-consistent counterparts. The results indicate that the proposed framework consistently improves classification performance and reduces variance across runs. Notably, representation-consistent models demonstrate enhanced stability under noisy input conditions, supporting the theoretical motivation of the proposed approach.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance Comparison on the MIT-BIH Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-score (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e90.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e90.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e91.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRC-LSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRC-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e93.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"7. Discussion","content":"\u003cp\u003eThe proposed framework introduces a principled inductive bias for medical time-series modeling by enforcing temporal [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] coherence in latent representations. Unlike architectural changes, the method is simple, interpretable, and broadly applicable.\u003c/p\u003e \u003cp\u003eThe approach also improves interpretability, as smoother hidden-state trajectories better reflect gradual physiological changes rather than noise-induced artifacts [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e"},{"header":"8. Limitations and Future Work","content":"\u003cp\u003eWhile effective, the proposed framework introduces an additional hyperparameter λ, which requires tuning. Future work will explore adaptive weighting strategies and extensions to transformer-based temporal models.\u003c/p\u003e"},{"header":"9. Conclusion","content":"\u003cp\u003eWe presented a representation-consistent gated recurrent framework for robust medical time-series classification. By explicitly constraining hidden-state evolution, the proposed approach improves stability, robustness, and generalization without altering existing gated architectures. This work highlights the importance of representation-level regularization for a reliable medical AI system\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735\u0026ndash;1780\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCho K, van Merri\u0026euml;nboer B, Bahdanau D, Bengio Y (2014) On the properties of neural machine translation: Encoder\u0026ndash;decoder approaches. arXiv preprint arXiv:1409 1259\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChung J, Gulcehre C, Cho K, Bengio Y (2014) Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412 3555\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChe Z, Purushotham S, Cho K, Sontag D, Liu Y (2018) Recurrent neural networks for multivariate time series with missing values. Sci Rep, 8, 1\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajkomar A, Dean J, Kohane I (2019) Machine learning in medicine. N Engl J Med 380(14):1347\u0026ndash;1358\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLipton Y, Kale D, Elkan C, Wetzel R Learning to diagnose with LSTM recurrent neural networks. arXiv preprint arXiv:1511.03677, 2015.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsteban C, Hyland SL, R\u0026auml;tsch G (2017) Real-valued (medical) time series generation with recurrent conditional GANs. arXiv preprint arXiv:1706.02633\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrivastava J, Greff K, Schmidhuber J (2015) Training very deep networks. Adv Neural Inf Process Syst\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoffe S, Szegedy C (2015) Batch normalization: Accelerating deep network training by reducing internal covariate shift, International Conference on Machine Learning\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMikolov T, Karafi\u0026aacute;t M, Burget L, Černock\u0026yacute; J, Khudanpur S (2010) Recurrent neural network based language model, Interspeech\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGraves A (2012) Supervised sequence labelling with recurrent neural networks. Springer\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen RTQ, Rubanova Y, Bettencourt J, Duvenaud D (2018) Neural ordinary differential equations. Adv Neural Inf Process Syst\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGajbhiye N, Singh KK An Optimized Depression Detection Technique Using Behavioral Analysis and Machine Learning, in Proc. 2025 6th International Conference on Recent Advances in Information Technology (RAIT), Dhanbad, India, 2025, pp. 1\u0026ndash;5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/RAIT65068.2025.11089439\u003c/span\u003e\u003cspan address=\"10.1109/RAIT65068.2025.11089439\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGajbhiye N, Singh KK (2025) Meta Heuristic Based Optimized Intelligent Framework for Kidney Disease Detection Using Deep Learning, 2025 4th OPJU International Technology Conference (OTCON) on Smart Computing for Innovation and Advancement in Industry 5.0, Raigarh, India, pp. 1\u0026ndash;5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/OTCON65728.2025.11070857\u003c/span\u003e\u003cspan address=\"10.1109/OTCON65728.2025.11070857\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh KK, Gajbhiye N, Mishra GS (2025) Exploring Multi-Stage Deep Convolutional Neural Network for Medicinal Plant Disease Diagnosis, in Proc. 6th International Conference on Deep Learning, Artificial Intelligence and Robotics (ICDLAIR 2024), Atlantis Press, pp. 87\u0026ndash;101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2991/978-94-6463-740-3_9\u003c/span\u003e\u003cspan address=\"10.2991/978-94-6463-740-3_9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGajbhiye N, Singh KK, Mishra GS (2025) Enhancing Crop Disease Detection Systems with Explainable AI Techniques for Deep Learning Models Using Spectral Imaging, in Proc. 6th International Conference on Deep Learning, Artificial Intelligence and Robotics (ICDLAIR 2024), Atlantis Press, pp. 110\u0026ndash;126. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2991/978-94-6463-740-3_11\u003c/span\u003e\u003cspan address=\"10.2991/978-94-6463-740-3_11\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh KK (2024) A Machine Learning Model for Cerebral Palsy Disorder Detection in Integration with Hybrid Optimization, International Journal of Intelligent Systems and Applications in Engineering, vol. 12, no. 23s, p. 2550, Nov. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.17762/ijisae.v12i23s.7390\u003c/span\u003e\u003cspan address=\"10.17762/ijisae.v12i23s.7390\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumari S, Singh KK, Nand P, Mishra GS, Astya R (2023) A Comparative Study of Security Issues and Attacks on Underwater Sensor Network, in Proc. Third International Conference on Computing, Communications, and Cyber-Security, Lecture Notes in Networks and Systems, vol. 421, Springer, Singapore. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-981-19-1142-2_5\u003c/span\u003e\u003cspan address=\"10.1007/978-981-19-1142-2_5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh KK, Kumar S, Nand P, Mishra S A Comparative Study on Energy-Efficient and Non Energy-Efficient Routing Protocols in Underwater Sensor Networks, in Proc. 2021 5th International Conference on Information Systems and Computer Networks (ISCON), Mathura, India, 2021, pp. 1\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ISCON52037.2021.9702477\u003c/span\u003e\u003cspan address=\"10.1109/ISCON52037.2021.9702477\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIsmail H, Fawaz et al (2019) Deep learning for time series classification: A review. Data Min Knowl Disc 33(4):917\u0026ndash;963\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":true,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Sharda University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Recurrent Neural Networks, Medical Time-Series Analysis, Representation Learning, Temporal Consistency, Gated Recurrent Units, Healthcare Artificial Intelligence","lastPublishedDoi":"10.21203/rs.3.rs-8832658/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8832658/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMedical time-series data are characterized by irregular sampling, high noise levels, missing values, and strong inter-feature dependencies. Recurrent neural networks (RNNs), particularly gated architectures such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are widely used for modeling such data due to their ability to capture temporal dependencies. However, standard gated recurrent models do not explicitly constrain the evolution of latent representations over time, leading to representation drift and instability under noisy or incomplete inputs.\u003c/p\u003e \u003cp\u003eIn this work, we propose a representation-consistent gated recurrent framework (RC-GRF) that introduces a principled regularization strategy to enforce temporal consistency in hidden-state representations. The proposed framework is model-agnostic and can be integrated into existing gated recurrent architectures without modifying their internal gating mechanisms. We provide a theoretical analysis demonstrating how the consistency constraint bounds hidden-state divergence and improves stability. Extensive experiments on medical time-series classification benchmarks show that the proposed approach improves robustness, reduces variance, and enhances generalization performance, particularly in noisy and low-sample settings.\u003c/p\u003e","manuscriptTitle":"A Representation-Consistent Gated Recurrent Framework for Robust Medical Time-Series Classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-11 08:58:42","doi":"10.21203/rs.3.rs-8832658/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9059558e-65f2-4e11-83f1-d297c15d80ab","owner":[],"postedDate":"February 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62604911,"name":"Artificial Intelligence and Machine Learning"},{"id":62604912,"name":"Biomedical Engineering"},{"id":62604913,"name":"Medical Informatics"}],"tags":[],"updatedAt":"2026-02-11T08:58:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-11 08:58:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8832658","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8832658","identity":"rs-8832658","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}