Indonesian J our nal of Electrical Engineering and Computer Science V ol. 41, No. 2, February 2026, pp. 564 ∼ 578 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v41.i2.pp564-578 ❒ 564 J oint angle pr ediction and joint-type classication in human gait analysis using explainable deep r einf or cement lear ning Deepak N. R. 1,4 , Soumya Naik P . T . 2,4 , Ambika P . R. 2,4 , Shaik Say eed Ahamed 3,4 1 Department of Information Science and Engineering, Atria Institute of T echnology , Bang alore, India 2 Department of Computer Science and Engineering, City Engineering Colle ge, Bang alore, India 3 Department of Computer Science and Engineering (DS), Atria Institute of T echnology , Bang alore, India 4 V isv esv araya T echnological Uni v ersity , Belag a vi, India Article Inf o Article history: Recei v ed Mar 22, 2025 Re vised Oct 15, 2025 Accepted Jan 11, 2026 K eyw ords: Deep learning Explainable articial intelligence Human g ait analysis Maximization Q-learning and mutual information Rehabilitation Reinforcement learning ABSTRA CT Human g ait analysis is a k e y component of rehabilitation, prosthetics, and sports science, especially for clinical e v aluation and the de v elopment of adapti v e assis- ti v e technologies. Accurate joint-angle estimation and dependable joint-type classication remain dif cult because of the comple x temporal beha vior of g ait signals and the limited interpretability of man y deep learning (DL) approaches. While recent techniques ha v e enhanced predicti v e accurac y , their clinical appli- cability is often limite d by insuf cient transparenc y and adaptability in learning mechanisms. T o o v ercome these limitations, this w ork proposes an inte grated frame w ork that unies DL, reinforcement learning (RL), and e xplainable arti- cial intelligence (XAI). Stochastic depth neural netw orks (SDNN) are applied for joint-angle re gression, whereas deep feature f actorization netw orks (DFFN) are used for multi-class joint-type classication. Optimization is achie v ed using Q-learning (QL) and mutual information maximization (MIM), ensuring stable con v er gence and impro v ed learning ef cienc y . T o impro v e interpretability , the frame w ork incorporates counterf actual and contrasti v e e xplanations, feature ab- lation studies, and prediction proba bility analysis. Experimenta l ndings sho w that the SDNN MIM model attains an R 2 score of 0 . 9881 , with RL re w ards increasing from 0 . 997 to 0 . 999 during re gression training. F or joi nt-type clas- sication, the DFFN MIM model achie v es an accurac y of 0 . 95 , with re w ard v alues impro ving from 0 . 90 to 0 . 98 . These results demonst rate the ef fecti v e- ness of the proposed frame w ork in deli v ering accurate and interpretable g ait predictions, supporting its rele v ance to biomechanics, healthcare, personalized rehabilitation, and intelligent assisti v e systems. This is an open access article under the CC BY -SA license . Corresponding A uthor: Shaik Sayeed Ahamed Department of Computer Science and Engineering (DS), Atria Institute of T echnology Bang alore, Karnataka, 560064 India Email: shaik.sayeedahamed1999@gmail.com 1. INTR ODUCTION Human g ait analysis constitutes a core research domain in biomechanics, rehabilitation, prost hetics, and sports science, with signicant rele v ance to clinical diagnosis, rehabilitation e v aluation, and the de v el- opment of intelligent assisti v e technologies. Accurate g ait assessment supports early detection of mo v ement impairments, f acilitates impro v ed prosthetic and orthotic design, and aids in injury pre v ention. Con v entional g ait analysis approaches primarily rely on motion-capture systems, force plates, and handcrafted biomechanical J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 565 models. While ef fecti v e in controlled laboratory settings, these methods often f ace challenges related to high- dimensional data, inter -subject v ariability , limited generalization across di v erse mo v ement patterns, and e xten- si v e manual feature engineering requirements. In recent years, deep learning (DL) methods ha v e been widely adopted to address these limitations by automatically learning hierarchi cal representations from g ait data. De- spite their ef fecti v eness, DL–based g ait models still e xhibit notable limitations, including o v ertting, restricted interpretability , and inef cient optimization, which constrain their clinical reliability . T o mitig ate these issues, e xplainable articial intelligence (XAI) and deep reinforcement learning (DRL) ha v e g ained increasing atten- tion in g ait analysis research. In the conte xt of clinical g ait analysis (CGA), Slijepce vic et al. [1] cate gorized XAI techniques into data e xploration, predict ion e xplanation, and model e xplanation using approaches such as t-SNE and layer -wise rele v ance propag ation (LRP). Although these methods enhanced interpretability , re- inforcement learning (RL)–based optimization w as not e xplored. Lik e wise, Madanu et al . [2] emplo yed XAI for pain assessment, reducing subjecti vity b ut without capturing the sequential and biomechanical comple x- ity of g ait. SHAP-based e xplanation techniques reported in [3], [4] impro v ed cl inical condence; ho we v er , these studies were limited to supervised learning paradigms and lack ed adapti v e optimization strate gies. RL has sho wn strong potential in sequential decision-making and continuous control tasks. The soft actor -critic (SA C) frame w ork presented in [5], [6] enabled st able learning in continuous action spaces, and autonomous locomotion without predened motion models w as in v estig ated in [7]. These studies mainly addressed robotic locomotion, where biomechanical constraints, safety requirements, and interpretability dif fer from human g ait analysis. Guided SA C methods, such as [8], enhanced performance in partially observ able en vironments; ho w- e v er , limited polic y transparenc y restricts their clinical applicability . Model-based RL e xtensions incorporating uncertainty modeling and model predicti v e control (MPC) in [9] impro v ed sample ef cienc y , yet their rele v ance to human g ait remains constrained by safety and e xplainability concerns. Recent surv e ys and re vie ws [10]–[12] highlighted the promise of DRL for g ait analysis and rehabilitation while i dentifying ongoing challenges, in- cluding small clinical datasets, dependence on simulated en vironments, and limited interpretability of learned policies. Explainable RL taxonomies in [12] and roadmap studies in [13] further emphasized the dif culty of e xplaining sequential decision-making processes in safety-critical applications. IMU-based g ait in v estig ations such as [14] demonstrated ef fecti v e prediction of dynamic balance b ut did not incorporate reinforcement-dri v en optimization or biomechanical interpretability . Similarly , GRF-based g ait classication in [15] utilized SHAP- based e xplanations without adapti v e learning mechanisms. Be yond g ait-focused research, XAI applications in healthcare and sports analytics [15], [16] reported challenges related to dataset quality , predicti v e performance, and generalization. Ethical transparenc y and accountability in machine learning were emphasized in [17], while sensiti vity to dataset bias w as discussed in [18] and [19]. Recent XAI-enabled g ait decision-support studies [20], [21] applied LIME and SHAP to support clinical reasoning b ut encountered scalability and real-time in- terpretability limitations. Finally , [13] achie v ed strong foot-condition classication using handcrafted features and LIME e xplanations, yet lack ed automated feature learning and reinforcement-based optimization. Ov er - all, although prior studies demonstrate substantial progress in XAI and DRL for human mo v ement analysis, a unied frame w ork inte grating deep neural netw orks, RL–dri v en optimization, and e xplainable mechanisms for accurate, adapti v e, and clinically interpretable human g ait prediction remains insuf ciently in v estig ated. Despite the substantial progress achie v ed through deep learning in human g ait analysis, se v eral open challenges still restrict its clinical applicability . Most e xisti n g w orks depend on post-hoc interpretability meth- ods applied to supervised learning models, which pro vide limited insight into model beha vior and of fer minimal e xplanation of sequential decision-making processes. As a result, the inte gration of XAI within RL–based g ait analysis frame w orks remains lar gely undere xplored. Furthermore, current g ait modeling strate gies frequently f ace optimization challenges, including unstable training beha vior , limited adaptability to time-v arying g ait patterns, and reduced generalization acr o s s subjects and mo v ement conditions. Although RL approaches, such as SA C, ha v e demonstrated strong performance in roboti c locomotion, their ef fecti v eness for modeling human g ait dy na mics—particularly for combined re gression and multi-class classication tasks—has not been thor - oughly e xamined. Addressing these g aps, this study proposes a unied deep learning frame w ork augmented with RL and e xplainability components to enhance predicti v e accurac y , learning stability , and clinical inter - pretability in g ait analysis. F or joint-angle estimation, stochastic depth neural netw orks (SDNN) are adopted to impro v e generalization by dynamically bypassing netw ork layers during training. T o ensure stable and ef cient optimization, QL and MIM are inte grated into the learning process. F or joint-type classication, deep feature f actorization netw orks (DFFN) are emplo yed to der i v e discriminati v e spatio-temporal g ait representations, sup- porting rob ust multi-cl ass decision-making. In addition, adv anced XAI techniques—including counterf actual J oint angle pr ediction and joint-type classication in human gait analysis using e xplainable ... (Deepak N. R.) Evaluation Warning : The document was created with Spire.PDF for Python.
566 ❒ ISSN: 2502-4752 and contrasti v e e xplanations, feature ablation analysis, and prediction condence assessment—are incorporated to deli v er clinically meaningful insights and enhance trust in model predictions. Ov erall, this w ork contrib utes a RL–dri v en and e xplainable g ait analysis frame w ork that unies accurate prediction, adapti v e learning, and transparent decision-making. The proposed methodology establishes a basis for reliable g ait modeling applica- ble to intelligent assisti v e systems and future clinical deplo yment. The remainder of this paper is structured as follo ws: section 2 describes the dataset, preprocessing steps, problem formulation, model architectures, and the inte gration of RL and XAI strate gies, section 3 presents the e xperimental results and interpretability analysis, and section 4 concludes the study with clinical implications and future research directions. 2. METHOD 2.1. Resear ch design The increasing demand for data-dri v en and clinically dependable human mo v ement analysis high- lights the challenge of accurately modeling comple x g ait dynamics. This study concentrates on de v eloping a unied frame w ork capable of performing joint-angle re gression and multi-class joint-type classication while maintaining rob ustness, learning stability , and clinical interpretability . T o accomplish this, the proposed ap- proach inte grates deep neural architectures with RL and mutual information–based optimization, forming a cohesi v e pipeline illustrated in Figures 1–4. F or joint-angle estimation, SDNN are emplo yed to capture tem- poral joint trajectories. As sho wn in Figure 1, SDNN utilizes a probabilistic layer -skipping strate gy in which each netw ork block (P0–P3) is assigned a survi v al probabil ity . Shallo wer layers remain acti v e during training, while deeper layers are selecti v ely bypassed. When a layer is skipped, its output is substituted with a shortcut connection from the preceding layer , enabling uninterrupted forw ard propag ation. This architecture mitig ates o v ertting, enhances generalization, and promotes stable learning from noisy and v ariable g ait signals by learn- ing hierarchical temporal representations. F or mul ti-class joint-type classication, deep feature f actorization (DFF), depicted in Figure 2, is applied to enable structured feature decomposition and dimensionality reduc- tion. Ra w g ait signals are initially processed through feature e xtraction and reshaped into matrix form, which is subsequently f actorized into basis and acti v ation matrices. Methods such as singular v alue decomposition, non-ne g ati v e matrix f actorization, or principal component analysis produce compact yet informati v e represen- tations that preserv e essential spatio-temporal characteristics while reducing redundanc y , thereby impro ving both discriminati v e capabil ity and computational ef cienc y . T o support adapti v e optimization, RL is incor - porated through a QL mechanism, as illustrated in Figure 3. In this conguration, the model functions as an agent that recei v es re w ard feedback based on prediction performance. Incorrect predictions generate correc- ti v e re w ards, directing iterati v e Q-v alue updates and polic y renement. Through continuous interaction and feedback, the model progressi v ely impro v es learning stability and classication accurac y . Complementing this process, MIM, sho wn in Figure 4, is emplo yed to reinforce feature rele v ance across modalities. By maximizing shared information among complementary feature representations, MIM ensures that retained features remain informati v e and non-redundant, ultimately impro ving representation quality and do wnstream performance. Figure 1. Flo w diagram of SDNN Figure 2. Flo w diagram of DFF Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 2, February 2026: 564–578 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 567 2.2. Data sour ces and pr epr ocessing This study utilizes a multi v ariate human g ait dataset sourced from the UCI machine learning repos- itory , released on December 14, 2022. The dataset consists of 181,800 time-series samples acquired from 10 health y participants performi ng g ait under three e xperimental conditions: unbraced, knee-braced, and ankle- braced. Under each condition, participants completed 10 g ait c ycles, with joint-angle trajectories captured at 101 discrete time points corresponding to a complete g ait c ycle. Each sample is characterized by se v en at- trib utes, including subject identier , w alking condition, replicati on inde x, le g side, joint type (ankle, knee, or hip), time step, and joint angle e xpressed in de grees. Data acquisition w as conducted at the Human Dynamics and Controls Laboratory , Uni v ersity of Illinois at Urbana–Champaign [22]–[24], and the dataset contains no missing entries. The balanced distrib ution across subjects, w alking conditions, limbs, and joint types mak es the dataset appropriate for both re gression and classication ta sks in biomechanical g ait analysis. F or the joint-angle re gression task, the objecti v e w as to estimate continuous joint-angle v alues using subject-specic and g ait-related features. Data preprocessing in v olv ed loading the dataset with P andas, encoding cate gorical v ariables, and normalizing numerical features using MinMaxScaler . A feature matrix comprising 29 predic- tors w as formed, with joint angle designated as the re gression tar get. The dataset w as subsequently split into training and testing subsets and reshaped into sequential formats compatible with the SDNN-based re gression architecture. F or multi-class joint-type classication, the aim w as to identify joint cate gories using the same input attrib utes. Joint labels were one-hot encoded, numerical features were normalized, and a dat aset con- taining 27 input features w as constructed using the identical train–test split. The classication data were then arranged into s tructured sequences suitable for the DFFN-based architecture. Ov erall, these preprocessing pro- cedures produced clean, balanced, and well-or g anized datasets, establishing a reliable basis for accurate and interpretable g ait analysis across v arying w alking conditions. Figure 3. Flo w diagram of QL Figure 4. Flo w diagram of MIM 2.3. Model ar chitectur e and justication This study proposes a unied frame w ork that inte grates neural net w o r ks (NN), RL, and XAI to ad- dress joint-angle re gression and multi-class joint-type classication in human g ait analysis. The o v erall w ork- o w starts wi th dataset preparation, where noise and outliers are managed, cate gorical v ariables are encoded, and numerical features are normalized usi ng Min–Max s caling. The processed data are then partiti on e d into training and testing sets to enabl e balanced and unbiased e v aluation. F or joint-angle re gression, tw o v ariants of the SDNN are de v eloped. The SDNN QL model incorporates QL to support polic y-dri v en optimization dur - ing training, while the SDNN MIM model applies MIM to enhance feature representation and generalization performance. Both v ariants are designed to ef fecti v ely capture temporal g ait dynamics while m inimizing pre- diction error in joint-angle estimation. Re gression performance is assessed using mean squared error (MSE), mean absolute error (MAE), and the coef cient of determinati on ( R 2 ), complemented by residual and per - formance plots that assist in v alidating learning stability and predicti v e reliability . F or multi-class joint-type classication, tw o DFFN v ariants are utilized. The DFFN QL model inte grates QL to optimize action-selection beha vior during classication, whereas the DFFN MIM model emplo ys MIM to reinforce learned feature em- beddings. These models are trained to discriminate among ankle, knee, and hip joint cate gories. Classication performance is measured using accurac y , precision, recall, F1 score, and prediction probability distrib utions, with additional insights deri v ed from confusion matrices, R OC curv es, and precision–recall plots. T o enhance J oint angle pr ediction and joint-type classication in human gait analysis using e xplainable ... (Deepak N. R.) Evaluation Warning : The document was created with Spire.PDF for Python.
568 ❒ ISSN: 2502-4752 transparenc y and clinical interpretability , the frame w ork incorporates multiple XAI techniques. Counterf actual e xplanations identify minimal changes in input features required to modify predictions, while contrasti v e e x- planations highlight dif ferences between predicted outcomes and alternati v e classes. Feature ablation analysis e v aluates the contrib ution of indi vidual input v ariables, and prediction probability analysis demonstrates model condence across both re gression and classication tasks. These interpretability ndings are presented through visual and te xtual representations to support clear understanding of model decision-making. The complete architecture is sho wn in Figure 5, where Figure 5(a) ill ustrates the SDNN QL MIM re gression model and Fig- ure 5(b) displays the DFFN QL MIM classication model. (a) (b) Figure 5. Model architectures (a) SDNN QL MIM for re gression and (b) DFFN QL MIM for multi-class classication 2.4. P erf ormance metrics The proposed g ait analysis frame w ork is assessed using standard performance metrics suitable for both joint-angle re gression and multi-class joint-type classication. These metrics are selected to capture prediction accurac y , learning stability , and generalization capa b i lity , which are critical for clinically dependable e v aluation using the SDNN QL MIM and DFFN QL MIM models. F or joint-angle re gression, model performance is e v aluated using MSE, MAE, and the coef cient of determination ( R 2 ). MSE places greater emphasi s on lar ger discrepancies between predicted and actual joint-angle v alues, whereas MAE of fers a more intuiti v e measure of a v erage prediction error . The R 2 metric reects ho w ef fecti v ely the model e xplains v ariance in joint- angle data, enabling meaningful comparison across dif ferent re gression models and optimization strate gies. F or mul ti-class joint-type classication, e v aluation concentrates on accurac y , precision, recall, F1 score, and prediction probability distrib utions. Accurac y represents o v erall cl assication ef fecti v eness, while precision and recall characterize class-specic reliability and sensiti vity . The F1 score balances these measures to pro vide a unied performance indicator . T o further analyze class-le v el beha vior and decision boundaries, confusion matrices, recei v er operating characteristic (R OC) curv es, and precision–recall plots are utilized. Collecti v ely , these metrics pro vide a comprehensi v e e v aluation of t he rob ustness and ef fecti v eness of the proposed g ait prediction frame w ork. 2.5. Integration of XAI techniques The proposed fr ame w ork incorporates multiple XAI techniques t o impro v e transparenc y and con- dence in black-box learning models applied to human g ait analysis. When interpretability is needed, input data are preprocessed and forw arded through the trained model to obtain predictions. Counterf actual e xplanations are subsequently generated by identifying minimal and plausible modications in the input that result in dif fer - ent prediction outcomes, ensuring clinical rele v ance. In parallel, contrasti v e e xplanations are utilized to com- pare the predicted outcome with alternati v e scenarios, thereby emphasizing the k e y features that dri v e model decisions. T o further e xamine feature rele v ance, feature ablati on is performed by systematically remo ving or perturbing indi vidual input v ariables and analyzing the resulting v ariations in model outputs. This procedure enables a quantitati v e e v aluation of feature importance. In the multi-class classication setting, prediction prob- ability analysis is applied to assess class-wise condence le v els and determine the features that most strongly Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 2, February 2026: 564–578 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 569 inuence the predicted joint cate gory . F or instance, when a sample is classied as Joint Class 2, the associated probability scores reect the relati v e contrib ution of the corresponding input features ( X v ariables). Collec- ti v ely , these XAI techniques deli v er clear and actionable insights into model beha vior . When combined with RL–based decision renement and mutual information–guided feature optimization, the frame w ork enables accurate, interpretable, and clinically meaningful joint-angle prediction and joint-type classication. 2.6. Real-w orld implications The proposed frame w ork, inte grating deep learning with RL and XAI, demonstrates strong pract ical rele v ance for biomechanics, rehabilitation engineering, prosthetics, and CGA. Accurate joint-angle prediction and joint-type classication can support clinicians in the early detection of mo v ement disorders, enable person- alized rehabilitation strate gies, and contrib ute to the design of more ef fecti v e prosthetic and assisti v e de vices. The incorporation of RL allo ws the models to adapt to e v olving g ait patterns and sustain stable performance across v arying w alking conditions. Moreo v er , the inclusion of XAI techniques—such as counterf actual and contrasti v e e xplanations, feature ablation, and prediction probability analysis—enhances transparenc y by en- abling clinici ans and domain e xperts to interpret and v alidate model predictions. This le v el of interpretability addresses common concerns relate d to black-box learning models and promotes responsible clinical deplo y- ment. By unifying adapti v e learning with e xplainable decision-making, the proposed frame w ork pro vides a practical basis for implementing intelligent g ait analysis systems in real-w orld en vironments. As data-dri v en human mo v ement analys is continues to adv ance, such adapti v e and e xplainable approaches are e xpected to play a signicant role in the de v elopment of assisti v e technologies and e vidence-based healthcare solutions. 2.7. Mathematical f ormulation This section pres ents concise mathematical formulations of the XAI techniques used in this study , namely counterf actual e xplanations, contrasti v e e xplanations, and feature ablation. These formulations de- scribe ho w minimal input perturbations inuence model predictions and enable transparent interpretation for both re gression and multi-class classication tasks. 2.8. Counterfactual explanations Counterf actual e xplanations identify the minimal modication to an input instance that changes the model’ s prediction. Input features are normalized using Min–Max scaling is dened as (1): x norm = x − x min x max − x min (1) The counterf actual objecti v e is dened by minimizing a loss function that shifts the prediction from the original output to a tar get outcome is dened as (2): L ( x ) = − P ( y tar get | x ) + P ( y orig | x ) (2) The optimal counterf actual instance is obtained as (3): x ∗ = arg min x L ( x ) (3) 2.9. Contrasti v e explanations Contrasti v e e xplanations analyze ho w small perturbations in the input alter the model’ s pre diction. A contrasti v e instance is generated by adding bounded Gaussian noise as (4): x con = clip ( x + N (0 , σ 2 ) , 0 , 1) (4) The model prediction for both original and perturbed inputs is gi v en by (5): ˆ y = f ( x ) (5) Dif ferences between these predictions highlight features that most strongly inuence decision boundaries. J oint angle pr ediction and joint-type classication in human gait analysis using e xplainable ... (Deepak N. R.) Evaluation Warning : The document was created with Spire.PDF for Python.
570 ❒ ISSN: 2502-4752 2.10. F eatur e ablation Feature ablation e v aluates the i mportance of indi vidual features by measuring prediction changes after feature remo v al. F or a gi v en feature j , the perturbed input is dened as (6): X ′ = X with X [: , j ] = 0 (6) The impact of the ablated feature is quantied by the absolute prediction dif ference as (7): L j = | f ( X ) − f ( X ′ ) | (7) T o enable f air comparison across features, the ablation scores are normalized as (8): L norm j = L j q P n j =1 L 2 j (8) Higher normalized scores indicate greater inuence of the corresponding feature on the model’ s output. 2.11. Hyper parameter tuning strategy Hyperparameter tuning w as conducted independently for the joint-angle re gression and multi -class joint-type classicati on tasks to ensure stable con v er gence and dependable model performance. F or the re gres- sion task, the SDNN model w as trained using a test split of 0.3 and a x ed random seed of 42 to guarantee reproducibility . The netw ork architecture comprised v e stochastic depth layers with a survi v al probability of 0.8. Each hidden layer included 32 neurons with ReLU acti v a tion, while a linear acti v ation function w as emplo yed at the output layer to enable continuous joint-angle prediction. Model optimization w as carried out using the Adam optimizer , which supported training stability and reduced o v ertting. F or the multi-class joint- type classication task, the DFFN model dened tar get v ariables as features be ginning with joint , wit h 30% of the dataset allocated for tes ting and the same random seed of 42. The architecture incorporated a feature f actorization layer with 512 neurons, follo wed by interaction layers consisting of 256, 128, and 64 neurons. Additional non-linear transformation layers with 128 and 64 neurons were included, and a dropout rate of 0.4 w as applied t o enhance generalization. The nal softmax layer contained three neurons corresponding to the joint-type classes. T raining w as performed using the Adam optimizer with an initial learning rate of 0.0001, e xponential decay steps of 10,000, a decay rate of 0.8, staircase decay enabled, and cate gorical cross-entrop y as the loss function to ensure stable and reliable classication. T ables 1 and 2 summarize the e xperime n t al congurations applied for the re gression and multi-c lass classication tasks, respecti v ely . Across all e xperiments, deep learning and RL parameters were maintained consistently to ensure f air comparison across dif ferent XAI techniques. T o support interpretability , XAI e xpla- nations were generated for both the initial and nal predictions. T able 1. RL parameter settings for re gression (QL vs. MIM) P arameters QL-re gression MIM-re gression T otal training epochs for RL model 30 30 Batch size for training 64 64 Initial e xploration rate ( ϵ ) 0.5 0.5 Exploration decay rate 0.99 0.99 Discount f actor ( γ ) 0.95 0.95 Frequenc y of updating tar get model 10 5 T ar get model for RL updates – Clone of main model Possible learning rate v alues [0.00001, 0.00005, 0.0001, 0.0005, 0.001] – Possible dropout rate v alues [0.2, 0.3, 0.3, 0.4, 0.5] – Possible action v alues – [(0.00001, 0.2, 0.6), (0.00005, 0.3, 0.7), (0.0001, 0.3, 0.8), (0.0005, 0.4, 0.9), (0.001, 0.5, 0.9)] Learning rate for Q-table updates 0.5 – Number of features in training set – X train . shape [1] Counter for successful episodes – 0 Re w ard function 1 / (1 + MSE ) 1 / (1 + MSE ) Maximum re w ard v alue 1.0 1.0 Re w ard threshold for success count 0.8 0.8 V erbosity le v el 0 0 Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 2, February 2026: 564–578 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 571 T able 2. RL parameter settings for multi-class classication (QL vs. MIM) P arameters QL-multi class MIM-multi class Number of training epochs 50 50 Batch size for training 64 64 Initial e xploration rate ( ϵ ) 0.9 0.9 Exploration decay rate 0.98 0.98 Discount f actor ( γ ) 0.99 0.99 Frequenc y of updating tar get model 10 epochs 10 Possible action v alues [(0.00001, 0.3, 128), (0.00005, 0.4, 256), (0.0001, 0.4, 512), (0.0005, 0.5, 1024), (0.001, 0.6, 2048)] [(0.00001, 0.3, 128), (0.00005, 0.4, 256), (0.0001, 0.4, 512), (0.0005, 0.5, 1024), (0.001, 0.6, 2048)] Learning rate for Q-table updates 0.9 – Number of features in dataset – X train . shape [1] Scaling f actor for intrinsic re w ard – 0.5 Dropout rate in hidden layers – 0.4 Number of neurons in interaction layers – 128, 256, 512, 1024, 2048 3. RESUL TS AND DISCUSSION 3.1. Experimental setup The e xperimental setup utilizes adv anced DL, RL, and XAI techniques to support ef cient and rob ust g ait analysis. Data preprocessing and performance e v aluation were performed using the scikit-learn library , while deep neural netw ork architectures were designed and trained with T ensorFlo w/K eras. RL components were incorporated to enable adapti v e optimizat ion during trai ning, and XAI techni qu e s were int e gra ted to impro v e transparenc y and interpretability . This unied setup f acilitates reliable joint-angle re gression and multi-class joint-type classication with clinically meaningful insights. 3.2. Exploratory data analysis and featur e insights Figure 6 illustrates a lollipop chart summarizing the mean v alues of al l input features. The time fea- ture sho ws the highest mean v alue (approximately 50 ), follo wed by the angle feature (approximately 12 . 15 ), indicating t heir dominant numerical magnitude within the dataset. In contrast, features such as subject, con- dition, replication, le g, and joint e xhibit lo wer mean v alues (ranging from 1 to 5 ), reecting their cate gorical or discrete nature. Figure 7 presents a line plot with error bars representing the mean and standard de viation of each feature. The time feature demonstrates both the highest mean and the greatest v ariability , whereas angle sho ws moderate v ariation. The remaining features display shorter error bars, indicating limited v ari- ability consistent with cate gorical attrib utes. Figure 8 depicts a he xbin plot visualizing the joint distrib ution of Class and Hypertension, where color intensity denotes data density . This visualization emphasizes dom- inant class–h ypertension combinations while minimizing visual clutter from indi vidual data points. Finally , the correlation matrix in Figure 9 indicates generally weak linear relationships among features, with a modest positi v e correlation ( 0 . 22 ) identied between time and angle. The o v erall lo w linear dependenc y supports the application of nonlinear and multi v ariate modeling approaches to capture comple x g ait dynamics. Figure 6. Lollipop chart of feature means Figure 7. Mean and standard de viation for each feature J oint angle pr ediction and joint-type classication in human gait analysis using e xplainable ... (Deepak N. R.) Evaluation Warning : The document was created with Spire.PDF for Python.
572 ❒ ISSN: 2502-4752 Figure 8. He xbin plot Figure 9. Correlation matrix 3.3. Regr ession perf ormance analysis Figure 10 pro vides a comparati v e e v aluation of inte grated NN and RL-based re gression models, where the SDNN frame w ork is optimized using QL and MIM. In Figure 10(a), the QL–based model displays a gradual rise in re w ard v alues from approximately 0.992 to 0.998 o v er 30 epochs, indicating steady performance im- pro v ement with minor uctuations. In contrast, Figure 10(b) illustrates that the MIM-based model con v er ges more quickly , increasing from about 0.997 to nearly 0.999 within the same epoch range. Ov erall, although both optimization strate gies demonstrate ef fecti v e learning beha vior , SDNN MIM achie v es f aster con v er gence and mar ginally higher re w ard v alues than SDNN QL, indicating superior optimization ef cienc y for joint-angle re gression tasks. Figure 11 presents a comparati v e assessment of inte grated NN and RL-based re gression models for joint-angle prediction, specically SDNN QL and SDNN MIM. Performance is e v aluated using MSE, MAE, and R 2 . The SDNN MIM model records lo wer errors (MSE = 0 . 0003 , MAE = 0 . 0125 ) compared to SDNN QL (MSE = 0 . 0006 , MAE = 0 . 0183 ) and achie v es a higher R 2 score ( 0 . 9881 vs. 0 . 9750 ), reecting impro v ed v ari- ance e xplanation and model t. These ndings suggest that MIM st rengthens feature learning and re gression accurac y , whereas QL is relati v ely less ef fecti v e. Ov erall, SDNN MIM emer ges as the most ef fecti v e model for joint-angle re gression, while maintaining strong interpretability . (a) (b) Figure 10. Model performance analysis (a) SDNN QL MIM for re gression and (b) SDNN QL MIM for re gression Figure 11. Comparati v e analysis of combined NN and RL-based re gression models Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 2, February 2026: 564–578 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 573 Figure 12 inte grates re gression performance analysis with XAI-based e xplanations. In Figure 12(a), counterf actual e xplanations of the SDNN QL MIM model analyze feature contrib utions across the rst, sec- ond, last, and last-to-rst predictions under QL and MIM. F or QL, the initial prediction is mainly dri v en by time , joint 2 , le g 2 , r eplication 1 , condition 3 , and subject 9 , while the nal prediction shifts to w ard joint 3 , r eplication 3 , condition 3 , and subject 4 . Under MIM, the second prediction emphasizes time , joint 3 , le g 2 , r eplication 8 , condition 3 , and subject 10 , whereas the last-to-rst prediction highlights time , joint 2 , le g 1 , r eplication 4 , condition 3 , and subject 9 . Across all predictions, time and condition 3 consistently emer ge as dominant features. Figure 12(b) presents contrasti v e e xplanations that further e xamine feature v ariations across prediction stages. F or QL, the rst prediction is inuenced by time , joint 2 , le g 2 , r eplication 1 , condition 3 , and subject 9 , while the nal prediction shifts to w ard joint 3 , r eplication 3 , condition 3 , and subject 4 . Under MIM, the second prediction highlights time , joint 3 , le g 2 , r eplication 8 , condition 3 , and subject 10 , whereas the last-to-rst prediction emphasizes time , joint 2 , le g 1 , r eplication 4 , condition 3 , and subject 9 . These nd- ings indicate stable temporal and condition-related features, with other v ariables adapting based on the learning strate gy . In Figure 12(c), feature ablation analysis assesses feature importance through sensiti vity comparisons across predictions. F or QL, the initial prediction is af fected by time , joint 2 , le g 2 , r eplication 1 , condition 3 , and subject 9 , while the nal prediction shifts to w ard joint 3 , r eplication 3 , condition 3 , and subject 4 . F or MIM, the second prediction is inuenced by time , joint 3 , le g 2 , r eplication 8 , condition 3 , and subject 10 , whereas the last-to-rst predicti on highlights time , joint 2 , le g 1 , r eplication 4 , condition 3 , and subject 9 . Across all XAI techniques , time and condition 3 consistently e mer ge as the most dominant and stable fea- tures inuencing the tar get v ariable ( angle ). Ov erall, temporal and condition-related f actors go v ern prediction stability , while joint, le g, replication, and subject identiers contrib ute adapti v ely to model renement in g ait joint-angle re gression. (a) (b) (c) Figure 12. Re gression performance analysis with XAI-based e xplanations: (a) counterf actual e xplanations, (b) contrasti v e e xplanations, and (c) feature ablation 3.4. Multi-class classication perf ormance analysis Figure 13 presents a comparati v e analysis of inte grated NN- and RL-based multi-class classicat ion models using DFFN optimized with QL and MIM across 50 epochs. In Figure 13(a), the QL–based model e xhibits a gradual and oscil latory increase in re w ard v alues from approximately 0 . 70 to 0 . 96 , indicating slo wer and less stable con v er gence. In contr ast, Figure 13(b) sho ws that the MIM-based model rapidly e xceeds 0 . 90 within the rst 10 epochs and stabilizes around 0 . 98 by epoch 50. Ov erall, while both optimization strate gies demonstrate ef fecti v e learning beha vior , MIM achie v es f aster con v er gence and greater learning stability , mak- ing it a more ef cient optimization strate gy than QL for multi-class joint-type classication. (a) (b) Figure 13. Model performance analysis: (a) DFNN QL MIM for multi-class classication and (b) DFNN QL MIM for multi-class classication J oint angle pr ediction and joint-type classication in human gait analysis using e xplainable ... (Deepak N. R.) Evaluation Warning : The document was created with Spire.PDF for Python.