Indonesian J our nal of Electrical Engineering and Computer Science V ol. 40, No. 2, No v ember 2025, pp. 1087 ∼ 1099 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v40.i2.pp1087-1099 ❒ 1087 Inter pr etable federated deep lear ning models f or pr edicting gait dynamics in biomechanics Shaik Say eed Ahamed 1 , Akram P asha 2 , Sy ed Ziaur Rahman 3 , D . N. Puneeth K umar 4 1 Department of CSDS, Atria Institute of T echnology , Bang alore, India 2 Department of Information T echnology , Colle ge of Engineering and T echnology , Fujairah Uni v ersity , Fujairah, United Arab Emirates 3 Department of Computer Science, Colle ge of Art, Science, IT and Communication, Uni v ersity of Kalba, Kalba, United Arab Emirates 4 Department of Mathematics, School of Engineering and T echnology , CMR Uni v ersity , Bang alore, India Article Inf o Article history: Recei v ed No v 24, 2024 Re vised Jul 9, 2025 Accepted Oct 15, 2025 K eyw ords: Biomechanics Deep neural netw orks Federated learning Gait dynamics prediction Interpretable articial intelligence LIME Pri v ac y-preserving machine learning ABSTRA CT Accurate prediction of human joint angle dynamics and reliable g ait classica- tion are essential for applications in rehabilitation, biomechanics, and clinical monitoring. T raditional machine le arning (ML) models trained on centralized data raise concerns about pri v ac y , scalability , and transparenc y . This study proposes a federated deep learning (DL) frame w ork that inte grates pri v ac y- preserving model training wi th interpretable predictions. Specically , a g ated recurrent unit - deep neural netw ork (GR U-DNN) h ybrid model is de v eloped for re gression of joint angles, while a Long short-term memory - con v olutional neural netw ork (LSTM-CNN) h ybrid model is designed for binary and multi- class g ait classication. The frame w ork is deplo yed using the federated a v- eraging (FedA vg) algorithm across simulated clients, with each client training locally on its data. T o enhance interpretability , the local interpretable model- agnostic e xplanations (LIME) algorithm is inte grated at the client le v el to gener - ate human-understandable e xplanations for model predictions. The e xperimen- tal results demonstrate signicant impro v ements, including a reduction in global mean squared error (GMSE) from 56 . 16 to 3 . 31 and an increase in R-squared score from 0 . 80 to 0 . 99 for re gression, along with classication accuracies of 0 . 97 (binary) and 0 . 94 (multi-class). This scalable, pri v ac y-preserving frame- w ork bridges the g ap between accurac y and transparenc y , of fering impactful applications in biomechanics, healthcare, and personalized medicine. This is an open access article under the CC BY -SA license . Corresponding A uthor: Akram P asha Department of Information T echnology , Colle ge of Engineering and T echnology , Fujairah Uni v ersity Fujairah, United Arab Emirates Email: a.pasha@fu.ac.ae 1. INTR ODUCTION Human g ai t dynamics, patterns of mo v ement during locomoti on, is of critical importance in bi ome- chanics, rehabilitation, and healthcare. Accurate prediction of joint angle dynamics enables personalized reha- bilitation strate gies and early diagnosis of g ait abnormalit ies. Ho we v er , traditional centralized machine learning (ML) models, while ef fecti v e, often struggle with issues related to data pri v ac y , scalability , and interpretability , particularly in sensiti v e en vironments such as healthcare settings [1]–[3]. These challenges require inno v ati v e solutions that can balance high predicti v e accurac y with pri v ac y preserv ation and transparent model beha vior . In response to these li mitations, federated learning (FL) has emer ged as a promising alternati v e. FL enables collaborati v e training of ML models on decentralized client de vices while k eeping data locally stored, thereby J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1088 ❒ ISSN: 2502-4752 enhancing data pri v ac y and system scalability [4]–[6]. FL has demonstrated potential in se v eral healthcare use cases, such as anomaly detection, disease prediction, and medical imaging [4], [7], [8]. Recent literature has also emphasized the gro wing need for interpretable FL frame w orks, e specially in sensiti v e domai n s lik e healthcare where transpare nc y , pri v ac y , and interoperability are critical. Haf f ar et al. [9] proposed an approach to e xplain predictions and adv ersarial attacks in FL settings using random forests, thereby im pro vi ng transparenc y and rob ustness in model outcomes across distrib uted nodes. Ros che witz et al. [10] introduced I FedA vg, a modication of the standard FedA vg algorithm that promotes interpretable data interoperability , f acilitating better alignment of local models in heterogeneous FL en vironments . Further , W ang applied Shaple y v alue theory to quantify feature contrib utions in FL, allo wing for transparent e v aluation of feature importance without compromising data pri v ac y [11]. In the conte xt of electronic health records, Shi et al. [12] adv ocated for the design of interpretable deep learning models capable of deli v ering human- understandable predictions, further reinforcing the v alue of e xplainability in clinical applications. Furthermore, Sun and W u [13] proposed a rob ust frame w ork for classifying multi-source sensor data with strong transferabil- ity and lo w communication o v erhead, enabling inte gration into real-w orld clinical w orko ws. Similarly , Ng et al. [14] applied machine learning for acti vity recognition using g ait position data, highlighting the ef fecti v eness of supervised models under controlled conditions. T ogether , these studies highlight the promise of combining federated architectures with g ait-based acti vity recognition for pri v ac y-preserving human mo v ement analysis. Recent adv ances in FL ha v e demonstrated its potential i n v arious domains, including c ybersecurity , health- care, and internet of things (IoT). In the biomedical eld, Meqdad et al. [15] implemented FL with a g aussian multi v ariate aggre g ation module to ef ciently classify ele ctrocardiogram signals. Furthermore, Mouhni et al. [16] introduced a long short-term memory (LSTM)-based h ybrid FL model for stress detection in wearable de vices, impro ving both accurac y and user trust . Se v eral studies ha v e in v estig ated the application of DL mod- els such as recurrent neural netw orks (RNNs) and con v olutional neural netw orks (CNNs) in human mo v ement analysis, achie ving notable impro v ements in accurac y [1] [17], [18]. Ho we v er , these ef forts are often limited to centralized architectures, raising concerns about data vulnerability and reduced generalizability [8], [19], [20]. Although recent ef forts ha v e introduced FL into healthcare domains [6], [21], [22], the inte gration of interpretable deep neural netw orks (IDNNs) within FL setups to predict g ait dynamics remains undere xplored. Research g aps and research contrib utions: despite considerable adv ances in the predicti v e modeling of g ait dynamics, se v eral critical research g aps remain unaddressed. Most e xisting models f ail to tackle the chal- lenges of pri v ac y , scalability , and i nterpretability simultaneously . Although interpretability techniques such as LIME ha v e been widely adopted in centralized settings, their inte gration with FL models - particularly in the conte xt of g ait-related prediction tasks - remains limited. Furthermore, there is a notable absence of unied approaches that can ef fecti v ely handle both re gression tasks (such as predicting joint angles) and classication tasks (such as identifying g ait patterns) within a single federated frame w ork. T o address these limitations, the present study introduces a no v el federated DL frame w ork that inte grates pri v ac y-sensiti v e training with interpretable prediction capabilities. The frame w ork incorporates a g ated recurrent unit - deep neural netw or (GR U-DNN) h ybrid model for the re gression of joint angle dynamics, and a LSTM-CNN h ybrid model for performing both binary and multi-class g ait classication. These models are trained across distrib uted clients using the federated a v eraging (FedA vg) algorithm to enable collaborati v e learning without data centralization. T o impro v e interpretability at the client le v el, the LIME algorithm is embedded in the frame w ork to e xplain the predictions of the model in a human-understandable manner . The core contrib utions of this research are as fol- lo ws. First, it presents the de v elopment of an interpretable and pri v ac y-preserving FL frame w ork specically designed for g ait dynamics prediction. Second, it demonstrates the implementation of dual-purpose DL models, GR U-DNN and LSTM-CNN, that are capable of ef fecti v ely handling both re gression and classication tasks. Third, the proposed frame w ork is e xperimentally v ali d a ted, yielding substantial performance g ains including a reduction in global mean squared error (GMSE) from 56 . 16 to 3 . 31 , and an increase in global R-squared from 0 . 80 to 0 . 99 in re gression tasks, along with classication accuracies of 0 . 97 for binary and 0 . 94 for multi-class tasks. Finally , the inte gration of LIME ensures transparent and inter p r etable model beha vior , f acilitating trust and usability in clinical settings. The remainder of this paper is structured as follo ws. Section 2 presents the materials and methods used in this study , including problem formulation, data p r e-processing strate gies, model architectures, and FL setup. Section 3 pro vides a comprehensi v e discussion of the e xperimental results, co v er - ing performance e v aluation, communication cost analysis, interpretability insights, and comparisons between re gression and classication tasks. Section 4 concludes the paper by summarizing the k e y ndings, describing practical implications, and proposing directions for future research. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 1087–1099 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1089 2. METHOD 2.1. Resear ch design W ith the increasing demand for data-dri v en solutions, especially in human mo v ement analysis, t here is a critical need for predicti v e models that ensure both accurac y and pri v ac y . This research addresses the chal- lenge of predicting joint angle dynamics, an essentia l aspect of biomechanics and healthcare, by de v eloping models capable of capturing subtle v ariations in locomotion. T o meet these demands, an FL frame w ork with IDNNs is adopted. FL enables decentralized model training without sharing ra w data, aligning with health- care’ s pri v ac y requirements. IDNNs impro v e model transparenc y through interpretability tools such as LIME, pro viding insight into prediction rationale. The system empl o ys a client-serv er FL topology . Clients train deep models on pri v ate data; the central serv er aggre g ates parameters using the FedA vg algorithm. T w o neural ar - chitectures are used: a GR U-DNN h ybrid for joint angle re gression, and an LSTM-CNN h ybrid for binary and multi-class g ait classication. Figure 1 outlines the o v erall system w orko w , including client-s ide processing, aggre g ation, and LIME-based interpretability . A structured e v aluation strate gy compares FL-based models with centralized baselines, using statistical metrics and interpretability outputs to comprehensi v ely assess per - formance. Figure 1. Frame w ork of the proposed research w ork 2.2. Data sour ces and pr epr ocessing The dataset used comprises 181,800 instances across 7 attrib utes, of fering deep insights into human locomotion within the domain of life sciences [23]–[25]. It features a balanced distrib ution across v ariables such as subject, condi tion, replication, le g, joint, time, and angle, supporting a rob ust analysis of ph ysiological and biomechanical patterns. The consistenc y of the data, free of missing v alues, along with its sequential, multi v ariate, and time series structure, mak es it suitable for classication, re gression, and clustering tasks. F or the re gression task, the GR U model requires data sequences with time steps. Preprocessing be gins by loading the data via pandas, applying one-hot encoding for cate gorical features, and min-max scaling for numerical v alues. The reshaped ‘ngles’ are used as the ground truth, and the dataset is split into training and testing sets in accordance with FL principles. F or binary and multi-class classication tasks, angle v alues are thresholded or binned accordingly , then one-hot encoded. The dataset is reshaped into LSTM-compatible sequences, labels are constructed, and trai n-test splits are appli ed. These preprocessing steps ensure the data is well structured for FL-based LSTM model training, enabling rob ust classication across g ait cate gories. 2.3. Model ar chitectur e and justication This research aims to de v elop predicti v e models for human locomotion that balance pri v ac y pres erv a- tion with scalability - k e y challenges in healthcare applications. Centralized models often struggle with data Interpr etable feder ated deep learning models for pr edicting gait ... (Shaikh Sayeed Ahamed) Evaluation Warning : The document was created with Spire.PDF for Python.
1090 ❒ ISSN: 2502-4752 sensiti vity and scalability in a v ariety of sources. T o address this, we adopt FL, which decentralizes model training and ensures sensiti v e data remain local. DNNs, kno wn for their ability to learn comple x data patterns, form the base architecture of our predicti v e models. Their capacity to e xtract hierarchical features mak es them well suited for modeling joint angle dynamics. F or re gression, a GR U-DNN h ybrid is used, using GR Us to capture temporal dependencies in sequen- tial joint angle data, impro ving the model’ s ability to track dynamic patterns. F or classication, an LSTM-CNN architecture is emplo yed, combining LSTMs for temporal sequence learning with CNNs for spatial feature e x- traction. This design enables rob ust binary and multi-clas s g ait classication. The architecture of both models is sho wn in Figure 2, where Figure 2(a) presents the GR U-DNN model and Figure 2(b) illustrates the LSTM- CNN setup. (a) (b) Figure 2. Model architectures: (a) GR U-DNN for re gression and (b) LSTM-CNN for binary and multi-class classication 2.4. F ederated lear ning and optimization techniques setup Figure 3 illustrates the FL setup and compares the optimization strate gies adopted in this study . Figure 3(a) outlines the FL w orko w using the FedA vg algorithm, where clients train local models on pri v ate da ta and transmit updates to a central serv er for aggre g ation. This decentralized process preserv es data pri v ac y , supports heterogeneity , and scales ef ciently , a k e y for the healthcare and biomechanics domains. FL a v oids central bottlenecks and le v erages parallel client computation for inclusi v e and pri v ac y-preserving training. The FedA vg algorithm enables iterati v e renement of a global model while ensuring client data remain local. Whereas, Figure 3(b) compares the optimization methods used: Adam for GR U-DNN and RMSprop for LSTM-CNN. Adam combines momentum with adapti v e learning rates for ef cient parameter tuning during re gression tasks, while RMSprop stabilizes updates in classication models by adjusting learning rates based on recent gradients. The global loss curv es plotted across communication rounds sho w the ef cienc y of both optimizers in reducing training loss o v er time. These insights support the selection of optimizer congurations based on con v er gence beha vior and task-specic stability requirements. 2.5. P erf ormance metrics Model performance in re gression is e v aluated using GMSE and global R-square d ( R 2 ). GMSE quan- ties the a v erage prediction error in all samples, while R 2 reects the v ariance in joint angles e xplained by the Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 1087–1099 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1091 model. Monitoring these during communication rounds re v eals the renement and consistenc y of the model in predicting joint dynamics in a decentralized setup. F or binary and multi-class classication, global loss and global accurac y are emplo yed. Global loss, computed vi a binary or cate gorical cross-entrop y , indicates predic- tion error , while global accurac y measures correct classication rates. T racking these metrics throughout FL communication rounds pro vides insight into the model’ s reliability under non-IID conditions, v alidating the ef fecti v eness of the proposed FL-based g ait prediction frame w ork. (a) (b) Figure 3. Ov ervie w of the FL frame w ork and optimization strate gy (a) the FL setup emplo ying the FedA vg algorithm for decentralized training and model aggre g ation across multiple clients, and (b) the comparison of optimization methods (Adam for GR U DNN and RMSprop for LSTM CNN) based on global loss performance across communication rounds 2.6. Integration of LIME LIME enhances interpretability within the FL frame w ork by of fering transparent and model- agno s tic e xplanations for the predictions of GR U-DNN and LSTM-CNN. Locally approximates comple x models to e x- plain indi vidual predictions, highlighting the inuence of specic input features. F or re gression, LIME perturbs the features around the data points and quanti es their impact on the predicted joint angles. In classication, it re v eals the discriminati v e contrib ution of features in binary and multi-class decisions. These insights clarify wh y particular outputs are produced, promoting model transparenc y . Inte grating LIME not only demysties black-box decisions b ut also b uilds stak eholder trust, critical in the hea lthcare and biomechanics domains. The ability to understand the rationale of the model reinforces accountability , ethical compliance, and condence in FL-based predicti v e modeling. 2.7. Real-w orld implications The proposed FL frame w ork with interpretable DNNs and LIME of fers practical applications in biomechanics, healthcare, and medical diagnostics. Accurate prediction of joint angle dynamics helps in mus- cle beha vior research, injury pre v ention, and treatment personalization. The pri v ac y-preserving nature of FL Interpr etable feder ated deep learning models for pr edicting gait ... (Shaikh Sayeed Ahamed) Evaluation Warning : The document was created with Spire.PDF for Python.
1092 ❒ ISSN: 2502-4752 ensures ethical use of sensiti v e healthcare data. By addressing the limitations of centralized models, this ap- proach fosters scalable and secure ML deplo yment. The FedA vg algorithm supports distrib uted training, while LIME enhances transparenc y , allo wing stak eholders to understand and trust model decisions. As mo v ement analytics e v olv es, pri v ac y-preserving and interpretable FL models lik e ours are poised to dri v e the ne xt w a v e of inno v ation in healthcare and biomechanics. 2.8. Hyper parameter tuning strategy A manual grid search w as used to optimize h yperparameters, e xploring learning rates (0.0005, 0.001, 0.005), batch sizes (16 , 32 , 64) , and GR U/LSTM units (32 , 64 , 128) , with v ariations in dropout and dense layers. M odel selection w as based on GMSE for re gression and classication accurac y for binary and multi- class tasks. Early stopping ( patience = 3) w as applied to pre v ent o v ertting. Figure 4 illustrates the train- ing trends in all communication rounds, conrming stable con v er gence and loss minimization in the selected congurations. Figure 4. Model performance trends across communication rounds T able 1. Re gression model parameter settings P arameter 5 clients 10 clients 20 clients GR U layers 1 1 2 GR U units 16 32 64, 32 Dense layers 1 1 1 Dense units 1 1 1 Learning rate 0.001 0.001 0.001 Epochs 10 10 15 Batch size 32 32 32 Pri v ac y parameter 0.1 0.1 0.1 3. RESUL TS AND DISCUSSION 3.1. Experimental setup This study used adv anced ML tools to ensure ef cienc y and rob ustness. The scikit-learn library [26] w as used for preprocessing and classical ML tasks, while T ensorFlo w [27] enabled the de v elopment and scaling of DL architectures. LimeSurv e y [28] f acilitated structured surv e y-based data collection, impro ving input quality . T ables 1 and 2 summarize the e xperimental settings for re gression, binary , and multi-class cl assi- cation tasks, respecti v ely . Each conguration reects FL principles, m aintaining a x ed pri v ac y parameter ( E psil on = 0 . 1) across v arying numbers of clients (5 , 10 , and 20) . LIME e xplanations were generated e v ery 20 th communication round. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 1087–1099 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1093 3.2. Regr ession perf ormance analysis Figure 5 presents the performance of the re gression model using the GMSE and the global squared R as metrics across communication rounds. The upper plot sho ws a steady decline in GMSE, indicating a reduced prediction error through federated updates. The lo wer p l ot sho ws increasing global R-squared v alues, reecting enhanced e xplanatory po wer in modeling joint angle dynamics. These trends conrm the ef fecti v eness and adaptability of FL in impro ving re gression performance o v er time. T able 2. Classication model parameter settings for binary and multi-class tasks P arameter T ask 5 clients 10 clients 20 clients LSTM layers Binary 1 1 1 LSTM units Binary 32 64 64 Con v1D layers Binary 1 1 1 Con v1D lters Binary 32 64 64 Con v1D k ernel size Binary 3 3 3 Dense layers Binary 1 1 1 Dense units Binary 1 1 1 Learning rate Both 0.001 0.001 0.001 Epochs Binary 20 15 20 Batch size Binary 32 64 64 Number of clients Multi-class 5 10 20 LSTM units (global model) Multi-class 32 64 64 Con v1D lters (global model) Multi-class 16 32 32 Dense units (global model) Multi-class 16 32 64 Epochs (client training) Multi-class 10 15 15 Batch size (client training) Multi-class 32 64 64 Samples per round Multi-class 10 10 10 Pri v ac y parameter Both Epsilon = 0.1 Epsilon = 0.1 Epsilon = 0.1 Communication rounds Both 100 100 100 LIME e xplanation rounds Both Ev ery 20 th round Ev ery 20 th round Ev ery 20 th round Figure 5. Re gression performance analysis Figure 6 inte grates communication time, GMSE trends, and R-squared comparisons between dif ferent FL client congurations (5, 10, and 20). Figure 6(a) sho ws the communication time analysis, which is essential for assessing scalability as the number of clients gro ws. Figure 6(b) illustrates the reduction in GMSE with increased communication rounds and clients, conrming impro v ed predicti v e accurac y through distrib uted learning. Figure 6(c) depicts the corresponding gro wth in R-squared v alues, with nal s cores reaching 0.95, 0.98, and 0.99, respecti v ely , demonstrat ing the model’ s gro wing abili ty to capture joint angl e dynamics with high interpretability . 3.3. Classication perf ormance analysis Figure 7 outlines binary and multi-class classication trends, demonstrating consistent i mpro v em ents in both loss and accurac y . In the 100 th round, all congurations achie v ed 0 . 97 global accurac y in binary classication. Figure 8 combines the analysis of bina ry classication in three k e y aspects: Figure 8(a) commu- nication time, Figure 8(b) global accurac y , and Figure 8(c) global loss. Communication remained ef cient in all FL client congurations. Accurac y trends sho w high sustained performance, while global loss curv es reect stable con v er gence throughout training. Interpr etable feder ated deep learning models for pr edicting gait ... (Shaikh Sayeed Ahamed) Evaluation Warning : The document was created with Spire.PDF for Python.
1094 ❒ ISSN: 2502-4752 (a) (b) (c) Figure 6. Re gression performance trends across clients: (a) communication time, (b) GMSE, and (c) R-squared for 5, 10, and 20 FL clients Figure 7. Classication performance analysis Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 1087–1099 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1095 (a) (b) (c) Figure 8. Binary classication metrics across FL clients: (a) communication time, (b) accurac y , and (c) loss Figure 9 presents multi-class classication analysis, combining Figure 9(a) communication time, F ig- ure 9(b) global accurac y , and Figure 9(c) gl obal loss. Accurac y impro v ed steadily across congurations, reach- ing 0 . 95 with 20 clients. The decline in loss v alues, do wn to 0 . 1463 , conrms reliable performance and scala- bility of the model. LIME visualizations further v alidated interpretability . 3.4. Cr oss-comparison and discussion The FL model sho wed im pro v ement s in GMSE and R 2 for re gression and high p r ecision in classi- cation tasks. Communication analysis conrmed scalability , while LIME pro vided transparenc y in the pre- dictions. These trends emphasize FL adaptability to both re gression and classication, with practical benets for real-w orld deplo yment in sensiti v e domains. FL models sho w strong applicability in healthcare and biome- chanics. The y enable pri v ac y-preserving prediction of g ait dynamics, assist in diagnosing abnormalities, and support personalized treatments. LIME-enhanced interpretability increases trus t, making the solution ethical and suitable for clinical use. T able 3 sho ws that our frame w ork outperform s pre vious studies in GMSE and Interpr etable feder ated deep learning models for pr edicting gait ... (Shaikh Sayeed Ahamed) Evaluation Warning : The document was created with Spire.PDF for Python.
1096 ❒ ISSN: 2502-4752 classication accurac y while preserving pri v ac y . The proposed study better performs the re gression and clas- sication accurac y than e xisting w orks while preser v i ng pri v ac y . The us e of FL impro v es scalability and the inte gration of LIME impro v es the transparenc y of the model, consistent with the ndings in [19], [20]. Limitations of the current research w ork: despite strong results, generalizability may be limited by the demographic scope of the dataset. LIME results were h ypothetical and require empirical v alidation. Com- munication o v erhead and con v er gence challenges remain in FL settings. LIME’ s local nature m ay miss global patterns; future w ork may e xplore SHAP or inte grated gradients. Adapti v e FL strate gies could further impro v e con v er gence and personalization in heterogeneous en vironments. (a) (b) (c) Figure 9. Multi-class classication metrics across FL clients: (a) communication time, (b) accurac y , and (c) loss Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 1087–1099 Evaluation Warning : The document was created with Spire.PDF for Python.