跳到正文
原文
Nature 健康护理主题·· 28 天前AI 评分27

集成无线sEMG框架与贝叶斯优化机器学习用于多任务运动分类与神经肌肉表征

An integrated wireless sEMG framework with Bayesian-optimized machine learning for multi-task locomotor classification and neuromuscular characterization

AI 导读

研究提出一种集成无线表面肌电信号(sEMG)采集与贝叶斯优化机器学习的框架,用于多任务运动分类与神经肌肉表征。该系统结合移动端应用实现实时数据可视化与远程监控,并采用多种归一化技术降低受试者间变异性。该框架旨在解决现有可穿戴设备缺乏标准化处理及真实世界适用性的问题,支持康复训练与假肢控制等应用。

正文

Introduction

Accurate discrimination of locomotor tasks is a fundamental challenge in wearable neuromuscular monitoring, with direct implications for rehabilitation assessment and the control of lower-limb assistive devices 1,2. Gait analysis enables precise identification of gait phases, one of the core tasks in movement assessment and supports the development of normative reference databases for diverse clinical applications 3,4. Among the available sensing modalities, surface electromyography (sEMG) is particularly advantageous because muscle activity is detected prior to the onset of movement and encodes neuromuscular control strategies, activation timing, and movement efficiency, making it an ideal modality for locomotor classification 5,6.

Traditional laboratory-based gait analysis systems, while offering high precision and comprehensive multi-channel recordings, impose critical constraints, including substantial cost, restricted measurement environments that fail to replicate real-world locomotion variability, and an inability to capture long-term ambulatory data during activities of daily living 7,8. This limitation has driven the broader evolution of wearable sensing technologies toward miniaturized, low-cost, and scalable platforms suitable for continuous, real-world physiological monitoring 9. The integration of wireless communication technologies with wearable sEMG platforms further represents a significant advancement in accessibility, facilitating real-time data visualization, remote monitoring, and seamless data collection across diverse daily activities 4.

For effective classification of multiple locomotor tasks—including level-ground walking, stair ascent, and ramp ascent—machine learning techniques have demonstrated strong potential for automated pattern recognition, with reported accuracy exceeding 95% under controlled laboratory conditions 1,10. Although complex, deep learning architectures have achieved comparably high accuracy, traditional machine learning with optimized features offers superior interpretability and deployability on resource-limited embedded systems. This advantage is particularly critical for wearable system integration, where computational efficiency, low power consumption, and real-time on-device processing are essential. Moreover, the interpretability of these models facilitates clinical validation, making them more suitable for real-world assistance and rehabilitation wearable applications 1,9.

Despite these advances, existing wearable sEMG-based systems face persistent challenges that limit widespread clinical adoption, most notably the absence of standardized normalization techniques across the majority of published studies and limited real-world applicability owing to predominantly laboratory-constrained evaluation protocols 5,11. To the best of our knowledge, no prior study has simultaneously addressed signal-normalization standardization, multi-task locomotion classification, multi-domain feature evaluation, and Bayesian-optimized machine learning classification within a single, integrated wearable sEMG framework—a gap that the present work directly addresses. Inspired by these observations, this study proposes a comprehensive system integrating (1) custom-designed wireless sEMG acquisition from lower extremity muscles with a mobile application interface, (2) advanced signal preprocessing incorporating multiple normalization techniques to reduce inter-subject variability 5, and (3) systematic machine learning-based classification employing Bayesian hyperparameter optimization to distinguish diverse gait patterns 12. The system features a custom-built multi-channel sEMG acquisition unit integrated with a mobile application for real-time data visualization and remote monitoring 13. The proposed framework provides a practical, accurate, and interpretable solution for continuous ambulatory monitoring, supporting rehabilitation training, prosthetic control, and human–computer interaction 1.

Related work

Gait recognition and locomotor task classification using sEMG have attracted substantial attention over the past decade, driven by advances in wearable sensing and machine learning. Building on these foundations, recent research has diversified in terms of hardware platforms, classification strategies, and signal-processing pipelines, as summarized in the comparative analysis presented in Table 1. Early studies demonstrated that sEMG signals recorded from lower limb muscles contain sufficient discriminative information to distinguish between level-ground walking, stair ascent, and ramp locomotion 1,2.

Table 1 Comparative summary of representative wearable sEMG and gait-analysis studies.

Full size table

Several recent studies have focused on wireless, wearable EMG platforms for real-world gait monitoring. Park et al. employed a commercial wireless EMG system (Delsys Trigno) with eight participants and showed that support vector machines (SVM) achieved classification accuracies of 96.4% and 95.2% for two- and three-class gait-phase discrimination, respectively, under 10-fold within-subject cross-validation 3. Yu et al. proposed a custom ESP32-based wireless sEMG system and used a convolutional neural network (CNN) to classify six lower-limb motion tasks, achieving a maximum accuracy of 97.96% during the training phase; however, the deep learning model introduced substantial computational demands and reduced interpretability 8. Hwang et al. integrated inertial measurement units (IMUs) with SVM, random forest, and XGBoost classifiers, reporting accuracy exceeding 91% across abnormal gait classification tasks with 10 healthy participants simulating joint impairments 9. Liu et al. developed a four-channel custom wireless sEMG system and, using MF, MDF, STD, and SampEn features, achieved Pearson correlation coefficients greater than 0.800 for 14 gait-parameter estimations, although the evaluation focused on young subjects at sarcopenia risk and omitted Bayesian hyperparameter optimization . Simonetti et al. combined a 64-channel EMG garment with five IMU sensors, obtaining an R2 of 0.65 ± 0.21 for ankle-torque estimation in four post-stroke participants 10. Gomez-Correa et al. designed a custom ESP32-based platform integrating sEMG, IMU, and FSR insoles, demonstrating a synchronization accuracy of 576 μs and data loss below 0.8% under indoor and outdoor technical validation 15. Rana and Dey applied detrended fluctuation analysis (DFA) to Noraxon sEMG and FSR signals from 87 participants, using statistical inference rather than classification 11. Zohirov et al. conducted a comprehensive classification study comparing five classifiers on two commercial EMG devices; their random-forest model achieved up to 98.7% accuracy with the Biosignalsplux device, declining to 91.2% and 89.3% under cross-subject validation 12.

Beyond the individual results summarized above, Table 1 reveals several shared methodological limitations. First, most reported performance figures derive from intra-subject or random-split evaluations such as 10-fold within-subject cross-validation 3, random-split training 8,10, or single-split testing which tend to overstate real-world generalizability, since data from the same subject can appear in both training and test sets. Only a minority of studies employ more rigorous subject-level testing, such as leave-one-session or cross-subject validation 10,15. Second, most systems omit standardized normalization prior to feature extraction, and none applies Bayesian hyperparameter optimization to classifier selection 10,15. Third, real-time ambulatory monitoring remains rare: only one reviewed platform was evaluated in both indoor and outdoor conditions, and only for technical validation rather than classification 14. Taken together, no prior work simultaneously integrates standardized normalization, multi-domain feature evaluation, and Bayesian-optimized classification under rigorous, subject-independent validation. The present work addresses this compound gap.

Contributions of this study

While prior work has addressed individual aspects of wearable sEMG acquisition, signal normalization, or feature optimization in isolation (Table 1), few studies integrate a custom low-cost platform with a systematically validated, multi-stage feature-reduction and classification pipeline. This study addresses that gap through the following contributions:

  • A custom, low-cost wireless sEMG acquisition platform. A lightweight, three-channel wireless system built around the ESP32 microcontroller, with real-time streaming to a companion mobile application. The system is designed for clinical deployability and ecological validity, enabling continuous ambulatory data collection outside laboratory settings 8,13.

  • A comprehensive signal-normalization framework. Three widely used approaches MVC normalization (MVCN), z-score normalization (ZSN), and min–max normalization (MMN) are systematically compared under a standardized MVC protocol, addressing a gap often underreported in the sEMG literature. This represents one of the first studies to comparatively assess normalization strategies within a multi-task locomotion classification pipeline 5,11.

  • A nested, genetic-algorithm-optimized feature selection strategy. A genetic algorithm embedded within nested cross-validation progressively reduces the feature space, followed by correlation-based pruning to eliminate redundancy. Performance is assessed under nested cross-validation, LOSO, and an independent held-out test set, with per-subject reporting to support reproducibility. The contribution of each feature domain to classification performance is systematically evaluated to identify the optimal feature set for each locomotor task, while maintaining model interpretability suitable for clinical translation 9,15.

  • An optimized machine-learning classification pipeline. Multiple classifiers are evaluated across time-, frequency-, and time–frequency-domain features, with hyperparameters tuned via Bayesian optimization. Bayesian hyperparameter optimization is employed to maximize classification performance across the three locomotor tasks — level walking, stair ascent, and ramp ascent.

  • A comparative benchmarking of traditional machine learning against deep-learning baselines. Classical models are evaluated alongside deep-learning baselines trained directly on raw sEMG signals, under identical training and evaluation protocols, to directly contrast their accuracy and computational cost. This analysis provides empirical justification for the design choices adopted in the present work and quantifies the added value of the integrated pipeline over existing approaches 1.

System architecture

The present study employed a custom-designed wearable sEMG acquisition system to capture myoelectric signals from three biomechanically critical lower-limb muscles—the tibialis anterior (TA), soleus (SOL), and gastrocnemius lateralis (GL)—during ecologically valid locomotor tasks, including level gait, stair ascent, and ramp ascent. These muscles were strategically selected for their complementary functional roles in ankle-joint control during locomotion: the TA serves as the primary dorsiflexor, while the soleus and gastrocnemius lateralis constitute the triceps surae complex, the principal plantar flexors responsible for propulsive force during push-off.

The complete system architecture comprises two primary subsystems: a hardware system responsible for signal acquisition and wireless transmission, and a software system for real-time visualization, data storage, and advanced offline analysis.

Hardware system

The hardware system is designed to acquire and transmit sEMG signals using a custom-developed wearable device, as illustrated in Fig. 1. The device integrates three muscle sensors, a microcontroller unit, a power supply module, and standardized electrode configurations to enable simultaneous multi-channel EMG monitoring in real time. Its modular design ensures robust data acquisition and seamless wireless communication with external devices via Bluetooth technology 4. The individual hardware components, including the muscle sensors, microcontroller unit, electrode configuration, power supply, and overall system integration, are described in detail in Sects. “Muscle sensors (sEMG Sensors)” through “System integration and cost” and are summarized in the system-level block diagram shown in Fig. 2.

Fig. 1

Wearable sEMG sensor-connection diagram.

Fig. 2

System Hardware Architecture

Muscle sensors (sEMG sensors)

Three muscle sensors are employed to detect the electrical activity generated by the targeted muscles. Each sensor performs differential amplification and analog conditioning of the raw sEMG signals before transmitting them to the microcontroller for digitization, effectively rejecting common-mode noise and motion artifacts through a high common-mode rejection ratio (> 80 dB). The sensors are specifically designed for integration with a microcontroller unit. The main technical specifications of the EMG sensors are presented in Table 2.

Table 2 Electrical and operating specifications of the muscle sEMG sensor modules.

Full size table

Microcontroller unit (ESP32).

The ESP32 microcontroller serves as the central processing unit, coordinating signal acquisition from all three muscle sensors and enabling wireless data transmission via Bluetooth 4 . In accordance with the Nyquist-Shannon sampling theorem and given the sEMG signal bandwidth of 20–500 Hz with dominant spectral energy between 50–150 Hz, a sampling rate of 1000 Hz was selected to prevent aliasing while maintaining compatibility with resource-constrained embedded processing. The microcontroller digitizes analog sEMG signals using integrated 12-bit analog-to-digital converters (ADC) at this sampling rate 4 1.

This on-device preprocessing approach minimizes wireless data transmission requirements and enables real-time assessment of signal quality before Bluetooth transmission. The ESP32 was selected for the technical specifications summarized in Table 3.​

Table 3 Technical specifications of the ESP32 microcontroller used for wireless acquisition.

Full size table

ESP32 was programmed to execute real-time signal preprocessing prior to Bluetooth transmission, leveraging its dual-core architecture for efficient parallel processing 16. Raw sEMG signals undergo a two-stage preprocessing pipeline executed by the microcontroller to enhance signal quality: bandpass filtering and full-wave rectification, as represented in the hardware architecture diagram in Fig. 2.

Bandpass Filtering: A second-order Butterworth bandpass filter (20–500 Hz) is implemented digitally on the ESP32 to remove baseline drift and high-frequency noise while preserving physiological signal components 6​

Full-Wave Rectification: The filtered signal is rectified on-chip to convert the bipolar waveform into a unidirectional representation, facilitating subsequent amplitude analysis 6​ 10 17.

Electrode configuration

A bipolar electrode configuration was employed for each muscle channel, with electrodes positioned according to SENIAM guidelines to ensure standardization across participants. Recording electrodes were placed 2 cm apart (center-to-center) along the muscle fiber direction to optimize signal acquisition while minimizing cross talk, with the reference electrode on electrically neutral tissue (e.g., medial malleolus or tibial tuberosity). Gelled Ag/AgCl surface electrodes (10 mm diameter) were used 7 18.

Power supply

The system supports two power configurations to accommodate different monitoring scenarios. For short-duration sessions, the device operates on an internally integrated rechargeable lithium polymer (LiPo) battery (3.7 V, 600 mAh), providing approximately 1.7 h of continuous operation (estimated based on a total current draw of 300 mA and 85% battery conversion efficiency). For extended full-day monitoring, an external rechargeable power bank with a 20,000 mAh capacity is connected via USB, providing up to 50 h of continuous operation (estimated based on a total current draw of 300 mA and 85% power bank conversion efficiency), thereby supporting prolonged monitoring sessions including extended walking and stair climbing tasks. In both configurations, an integrated battery management system ensures safe operation by preventing overcharging, over-discharging, and short-circuit conditions 8 .

System integration and cost

All hardware components including three sEMG sensor boards, the ESP32 microcontroller module, and the integrated LiPo battery are housed within a wearable Holter-style rigid enclosure (65 × 40 × 15 mm, total mass ≈ 100 g), with internal layout optimized to minimize electromagnetic interference between analog acquisition circuits and the digital processing unit (Fig. 2). A 7-pin DIN connector (per IEC 60,130–9 standard) on the lateral face interfaces with a multi-conductor shielded cable comprising six differential electrode wires, two per muscle channel and a single shared reference electrode, with shielding complementing the sensor-level common-mode rejection (> 80 dB). A toggle power switch enables straightforward on/off operation under field conditions. The Holter enclosure is secured to the lower limb using adjustable elastic straps, maintaining consistent electrode–skin contact geometry and minimizing movement-induced displacement artifacts throughout all locomotor tasks.

The novelty of the proposed hardware lies in the system-level realization of a compact, fully wearable, three-channel device at a fraction of the cost of comparable wearable EMG platforms. A component-level bill-of-materials analysis places the total cost of the complete three-channel system at approximately 10,000 EGP (~ 180 USD), representing roughly one order of magnitude below that of commercial research-grade multi-channel wireless EMG platforms 3,15,19. The reported analog bandwidth, ADC resolution, sampling rate, wireless protocol, and power values describe the implemented design and/or manufacturer-specified component characteristics; they should not be interpreted as results of a concurrent laboratory benchmark. In particular, the system’s noise floor and BLE packet-loss characteristics have not yet been empirically bench-characterized.

For the next hardware iteration, priority validation will include RMS-noise testing with short-circuited or resistive inputs after bandpass filtering, together with sequence-numbered packet-loss measurements at short, intermediate, and extended transmission distances. Until these measurements are completed, no experimentally verified comparative claim is made for CMRR, noise floor, input impedance, or packet loss.

Participant characteristics

A cohort of twenty-five healthy male volunteers participated in this investigation. The cohort was partitioned into two subsets: twenty participants were allocated to model training and cross-validation, while the remaining five were held out as a completely independent test set. Demographic and anthropometric characteristics are summarized in Table 4.

Table 4 Demographic and anthropometric characteristics of the study participants.

Full size table

The train/test partitioning was performed strictly at the participant level: the five test-set participants did not contribute any samples to the training/cross-validation set, and no participant appears in both sets. Following the same protocol (3 muscles × 3 tasks per participant), the training/cross-validation set comprised 180 muscle-level samples (20 participants × 3 muscles × 3 tasks), while the 45 muscle-level samples reported in “Final evaluation of GA-optimized classifiers” section originate exclusively from the five held-out participants (5 participants × 3 muscles × 3 tasks). No subject-level data leakage is therefore present. For classification, each participant-task condition contributed the averaged muscle-level observation obtained from the three recorded gait cycles; no gait cycle or derived sample from participants S21–S25 entered training or cross-validation.

All participants were physically active and free of musculoskeletal, neurological, or cardiovascular conditions affecting gait, as confirmed by medical screening. Recruitment was restricted to male participants to control for well-documented sex-related differences known to confound sEMG-based locomotor classification, including lower-limb muscle morphology and strength 20,25,33, subcutaneous adipose tissue thickness greater in females and known to attenuate sEMG amplitude and alter spectral content 16,21,25 and gait biomechanics such as hip and knee joint moments 17,22,23,32. Because these factors can influence volume-conductor properties between muscle fibers and recording electrodes, the male-only cohort was retained to reduce within-cohort anthropometric heterogeneity during this initial methodological validation. However, this design limits external validity and does not establish sex-robust or sex-inclusive generalization; validation in a larger sex-inclusive cohort is required before clinical deployment 18,19,22,23,24,25.

This study was approved by the Ethical Committee of the Faculty of Medicine, Suez University (Approval No. 146), and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment.

Experimental protocol

Participants performed three locomotor tasks — level walking, stair ascent, and ramp ascent — at a comfortable, self-selected pace, while surface EMG signals were recorded from the tibialis anterior (TA), soleus (SOL), and gastrocnemius lateralis (GL) muscles. Each gait cycle was segmented into three phases: heel strike, mid-stance, and toe-off (Fig. 3). Each participant completed three gait cycles per task, and the mean values across trials were used to ensure measurement reproducibility and signal consistency.

Fig. 3

Representative EMG activation patterns of the Tibialis Anterior (TA), Soleus (Sol), and Gastrocnemius Lateralis (GL) across the three phases of the gait cycle: heel strike, mid-stance, and toe-off. Arrows indicate the direction of force and progression during each phase.

Software system

The software system enables real-time data visualization, secure wireless transmission, and comprehensive offline signal processing. It comprises two primary components: a mobile application for data acquisition and monitoring and a MATLAB-based analytical pipeline for signal preprocessing, feature extraction, and machine learning classification.

Mobile application

A custom Android mobile application connects wirelessly to the wearable device via Bluetooth Low Energy (BLE) protocol, providing the following functionalities:

Real-Time Visualization: The application displays sEMG signals from three lower extremity muscles (tibialis anterior, gastrocnemius lateralis, and soleus) simultaneously, enabling immediate visual assessment of signal quality, electrode–skin contact adequacy, and muscle activation patterns during data acquisition as illustrated in Fig. 4.

Fig. 4

Real-time sEMG data acquisition and visualization system during (a) stair ascent and (b) ramp ascent.

Data Acquisition and Storage: All sEMG data are automatically recorded and stored in plaintext format on the smartphone’s internal storage, facilitating subsequent offline analysis and compatibility with standard signal processing tools 8 24.

Signal processing pipeline

Each processing stage is designed to enhance signal quality, extract meaningful physiological information, and enable robust classification of muscle activation patterns across various gait modalities, including level walking, stair ascent, and ramp ascent.

As described in “Microcontroller unit (ESP32)” section, bandpass filtering and full-wave rectification were performed on-device by the ESP32 microcontroller, ensuring that wirelessly transmitted signals are pre-conditioned and noise-reduced. Therefore, the software-based processing pipeline begins with normalization, followed by feature extraction, feature selection, and classification.​

Signal normalization

Signal normalization constitutes a critical preprocessing stage for standardizing sEMG amplitude variations across participants, recording sessions, and electrode placements, thereby ensuring valid inter-subject comparisons and optimal machine learning model performance 25 . The raw sEMG signal amplitude is inherently influenced by inter-individual differences in muscle mass, subcutaneous tissue thickness, electrode–skin impedance, and electrode positioning, necessitating the application of rigorous normalization procedures prior to feature extraction and classification. Based on a comprehensive evaluation of the existing literature, three normalization approaches with demonstrated efficacy for gait analysis and locomotion classification were identified and implemented: Maximum Voluntary Contraction normalization (MVCN), Z-score normalization (ZSN), and Min–Max normalization (MMN) 23. A comparative summary of these methods, including their respective mathematical formulations and variable definitions, is presented in Table 5.

Table 5 Summary of signal normalization methods.

Full size table

A standardized isometric MVC was recorded for each muscle with participants seated. Each trial consisted of a 30-s baseline rest period, a 5-s maximal-effort contraction cued by a timed auditory signal, and a 30-s recovery period; this sequence was repeated three times per muscle. The mean of the three trials was retained as the reference value. Electrode placement followed SENIAM guidelines 34 and was identical to that used during task recordings. This MVC reference value was subsequently used to compute the MVC-normalized (MVCN) amplitude values.

Feature extraction

Feature extraction is a critical preprocessing stage in which physiologically relevant and mathematically robust descriptors are derived from the rectified sEMG signals to enable effective pattern recognition and classification. These features are accordingly adopted in the present framework to balance discriminative power with computational efficiency suitable for embedded deployment 1,6. A total of 40 features were systematically extracted from each muscle channel, yielding 120 feature vectors per subject across the three target muscles (tibialis anterior, soleus, and gastrocnemius lateralis). These features were categorized into three computational domains:

  • Time-domain features (28 total): amplitude-based measures (MAV, RMS, IEMG), variability measures (SD, MAD, COV), crossing features (ZC, SSC, WA), complexity indices (KURT, SKEW, LTKEO), and the myopulse percentage rate (MYOP), which quantifies the proportion of signal amplitudes exceeding a predefined threshold.

  • Frequency-domain features (10 total): derived from the power spectral density (PSD) of the signal, including mean and median frequency (MNF, MDF), spectral moments (SM1, SM2), spectral entropy (SE), and related measures.

  • Time–frequency-domain features (2 total): wavelet packet energy (WPE) and discrete wavelet transform energy (DWT-db4), obtained using a fourth-order Daubechies wavelet to capture both the temporal and spectral dynamics of the non-stationary sEMG signal. 3

The mathematical formulation and variable definitions for the complete set of extracted EMG features are provided in Appendix A, Table A1.

Feature selection

Feature selection is a critical preprocessing step for mitigating the curse of dimensionality inherent to high-dimensional sEMG feature sets, in which redundant or irrelevant descriptors can degrade classification performance, increase computational overhead, and reduce model interpretability. In this study, three distinct feature-selection strategies were systematically evaluated and compared: principal component analysis (PCA) applied independently, a genetic algorithm (GA) applied independently, and a hybrid PCA–GA sequential pipeline. This comparative framework enables a comprehensive assessment of the relative efficacy of unsupervised, variance-based dimensionality reduction (PCA); supervised evolutionary feature selection (GA); and their synergistic integration in optimizing classification performance while maintaining computational efficiency. The procedural framework and mathematical formulations underlying each approach are detailed in the following subsections.

Principal component analysis (PCA)

In the first experimental configuration, PCA was applied independently as an unsupervised, linear dimensionality-reduction technique to project the original feature space onto a lower-dimensional subspace that retains most of the data variance, thereby eliminating multicollinearity among correlated features. 12 PCA operates by identifying orthogonal principal components that capture the maximum proportion of variance in the data, with components accounting for 95% of the total variance typically retained for subsequent classification. 6 The sequential procedural steps and corresponding mathematical formulations are summarized in Table 6.

Table 6 Procedural steps and mathematical formulation of PCA-based feature reduction.

Full size table

Genetic algorithm (GA)

A genetic algorithm was applied as a supervised, population-based evolutionary optimization technique to identify the optimal discriminative feature subset directly from the original high-dimensional feature space. Unlike PCA, the GA employs a supervised search strategy that directly optimizes classification performance while penalizing feature redundancy, thereby reducing overfitting risk and computational overhead. Drawing on the biological principles of selection, crossover, and mutation, the GA iteratively evolves a population of binary-encoded candidate subsets, in which each bit encodes the inclusion or exclusion of a single feature, guided by a composite fitness function that balances classification accuracy against subset size. The procedural steps, mathematical formulations, and operator definitions are summarized in Table 7.

Table 7 Genetic-algorithm operators and procedural steps used for feature selection.

Full size table

Hybrid PCA–GA feature selection strategy

In the third configuration, a sequential PCA–GA pipeline was implemented to combine the complementary strengths of both methods. In the first stage, PCA reduced the high-dimensional feature space to a compact set of orthogonal principal components, eliminating redundancy and multicollinearity among the original features. In the second stage, the GA operated on this decorrelated subspace to identify the optimal subset of principal components that maximizes classification accuracy while minimizing feature-set cardinality. By substantially narrowing the combinatorial search space, this two-stage architecture improves GA convergence efficiency and offers a principled balance between dimensionality reduction, adaptive feature selection, and computational tractability for embedded sEMG classification (Figs. 5, 6, 7, 8 and 9).

Fig. 5

Top 29 features ranked by composite statistical importance score.

Fig. 6

(a) Effect size (η2) for the top 29 features, with Cohen’s large-effect threshold (η2 = 0.14) marked, and (b) Tukey HSD pairwise-significance heatmap.

Fig. 7

Electromyographic amplitude profiles for the tibialis anterior, soleus, and gastrocnemius lateralis during the gait cycle.

Fig. 8

EMG analysis during stair ascent.

Fig. 9

EMG analysis during ramp ascent.

Model selection

Model selection was a key preliminary step in building a robust sEMG-based locomotor-task classifier. Eight supervised learning algorithms were considered: logistic regression (LR), random forest (RF), support vector machine (SVM), decision tree (DT), k-nearest neighbors (KNN), naive Bayes (NB), neural network (NNET), and extreme gradient boosting (XGBoost)—spanning linear, probabilistic, distance-based, tree-based, kernel-based, and neural paradigms. This methodological diversity enabled a fair, representative comparison across classifier families and informed the selection of models for subsequent hyperparameter optimization and deployment.

Nested genetic algorithm for feature selection

To eliminate feature-selection bias, the genetic algorithm was implemented within a nested cross-validation framework. Candidate feature subsets were proposed and evaluated exclusively on each outer fold’s training partition, while generalization performance was assessed only on the corresponding held-out outer fold. Thus, the feature-selection search could not be informed, even indirectly, by data used for performance evaluation. 28 29

Each candidate chromosome was represented as a binary feature-inclusion vector z = (z₁, z₂, …, z_d), where zⱼ ∈ {0,1} indicates whether the jth feature was retained. The selected subset associated with chromosome z was defined as S(z) = {j : zⱼ = 1}, with cardinality |S(z)|= Σⱼ zⱼ.

The fitness function jointly rewarded predictive performance and compactness of the selected representation. For an inner cross-validation estimate A_CV(z) and a total feature dimension d, the objective was formulated as: 30

$$F(z) = \alpha A\_CV(z) + (1 - \alpha ) [1 - |S(z)|/d], 0 \le \alpha \le 1$$

(1)

Here, α controls the relative emphasis placed on classification accuracy versus feature-set compactness. For each outer fold k, A_CV(z) was computed only within the outer-training partition, and the selected chromosome z*ₖ was then evaluated once on the untouched outer-validation partition:

$$ A\_LOSO = \left( {1/K_{{\text{o}}} } \right)\sum\limits_{{\text{k = 1}}} {_{{\text{o}}}^{{\text{K}}} Acc\left( {C\left( {z{*}_{{\text{k}}} } \right),D_{{\text{k}}}^{{{\text{out}}}} } \right)} $$

(2)

In Eq. (2), Kₒ is the number of outer folds, C(z*ₖ) is the classifier trained using the features selected in outer fold k, and Dₖᵒᵘᵗ is the corresponding held-out participant or outer test partition. Features with zero variance in the relevant training partition were excluded before ranking and model fitting according to σⱼ2 = 0 ⇒ zⱼ = 0.

Correlation-based pruning

Pairwise Pearson correlation analysis was applied after nested feature selection to identify redundant descriptors. For two features xᵢ and xⱼ measured over n observations, the Pearson correlation coefficient was calculated as: 31

$$ \rho_{{{\text{ij}}}} = \sum\limits_{{\text{n}}} {\left( {x_{{{\text{in}}}} - \overline{x}_{i} } \right)\left( {x_{{{\text{jn}}}} - \overline{x}_{{\text{j}}} } \right)} /\left[ {\sqrt {\sum\limits_{{\text{n}}} {\left( {x_{{{\text{in}}}} - \overline{x}_{{\text{i}}} } \right)^{2} } } \sqrt {\sum\limits_{{\text{n}}} {\left( {x_{{{\text{jn}}}} - \overline{x}\mathop {\lim }\limits_{{x \to \infty {\text{j}}}} } \right)^{2} } } } \right] $$

(3)

In Eq. (3), xᵢₙ and xⱼₙ denote the values of features i and j for observation n, while x̄ᵢ and x̄ⱼ are their sample means. A pair was classified as redundant when |ρᵢⱼ| exceeded the prespecified threshold of 0.90:

$$ R{ } = { }\left\{ {\left( {i,j} \right):{ }\left| {\rho_{{{\text{ij}}}} } \right| > { }0.90} \right\} $$

(4)

For every redundant pair, one descriptor was removed according to the predefined retention rule, prioritizing the feature with the larger composite importance score and, where scores were comparable, the feature with the lower residual correlation to the rest of the candidate set. The retained representation was therefore defined as S_pruned = S_nested − R and was carried forward for independent evaluation. This procedure removes highly redundant descriptors while preserving complementary discriminative information. Because this pruning step was applied after the nested-GA selection, the resulting 17-feature comparison is treated as a post-selection sensitivity analysis; the 44-feature nested-GA result remains the primary fully nested subject-independent estimate unless pruning is re-estimated inside every outer training fold.

Deep-learning baseline architectures

To provide a transparent like-for-like comparison with the feature-based classifiers, two lightweight networks were trained end-to-end on raw, temporally aligned three-channel sEMG windows without handcrafted features. The 1D-CNN comprises three Conv1D blocks with 32, 64, and 128 filters, respectively, a kernel size of 3, batch normalization and max-pooling with stride 2 after each block, followed by flattening, a 64-unit dense layer with dropout (rate 0.5), and a final 3-unit softmax output layer. The network contains 154,307 trainable parameters and requires approximately 2,500,579 FLOPs per forward pass. The BiLSTM comprises two stacked bidirectional LSTM layers, with 64 units per direction in the first layer and 32 units per direction in the second layer, with dropout between layers, followed by a 32-unit dense layer and a final 3-unit softmax output layer. It contains 77,187 trainable parameters and requires approximately 9,080,291 FLOPs per forward pass. Both networks were evaluated using the same held-out 5-participant test set, nested 5 × 5 cross-validation, and LOSO protocol used for the traditional machine-learning pipeline.

Hyperparameter tuning

Hyperparameters for each candidate classifier were optimized using Bayesian optimization, which provides efficient exploration of high-dimensional search spaces while mitigating both underfitting and overfitting. Stratified k-fold cross-validation was employed throughout, with k evaluated over the range {2, 3, …, 10} and the optimal value selected independently per classifier based on the highest mean validation accuracy. The hyperparameter search spaces were defined as follows:

  • Logistic Regression: regularization strength C ∈ [0.001, 1000] and penalty type ∈ {L1, L2}; Support Vector Machine: C ∈ [0.01, 1000], kernel ∈ {linear, RBF, polynomial}, and γ ∈ [0.001, 100];

  • Random Forest: number of estimators ∈ [98, 498], minimum leaf size ∈ 1,17, and features per split ∈ [1, p], where p denotes the total number of input features;

  • XGBoost: number of estimators ∈ [98, 498], learning rate (η) ∈ [0.01, 0.3], maximum tree depth ∈ 3,9, and subsample ratio ∈ [0.5, 1.0], consistent with configurations reported in prior sEMG classification studies 32;

  • Decision Tree: splitting criterion ∈ {Gini impurity, entropy}, minimum leaf size ∈ 1,17, and maximum depth ∈ 1,18;

  • k-Nearest Neighbors: number of neighbors ∈ 1,18, distance metric ∈ {Euclidean, Manhattan, Minkowski}, and weighting scheme ∈ {uniform, inverse-distance};

  • Naïve Bayes: feature distribution ∈ {Gaussian, kernel density estimation} with smoothing bandwidth ∈ [0.01, 1.0];

  • Neural Network: a single hidden layer architecture was adopted given the relatively constrained dataset size, with the number of units ∈ [5, 98], training algorithm ∈ {Levenberg–Marquardt, scaled conjugate gradient}, learning rate ∈ [0.0001, 0.01], and maximum epochs ∈ [98, 998] The optimal configuration for each classifier was selected based on mean cross-validated accuracy across the three locomotor task classes.

Model performance evaluation

Classifier performance was assessed using accuracy, precision, recall, and F1-score computed from the confusion matrix. Accuracy reflects the overall proportion of correctly classified samples, precision measures the fraction of true positives among predicted positives, and recall quantifies the fraction of true positives among all actual positives; mathematical definitions are reported in Table 8.

Table 8 Definitions of the classification-performance metrics derived from the confusion matrix.

Full size table

Statistical analysis

Descriptive statistics are presented as mean ± standard error of the mean (SEM) together with the coefficient of variation (CV%). Normality was assessed using the Shapiro–Wilk test, and homogeneity of variance was examined using Bartlett’s test and the Brown-Forsythe test before inferential analysis. Task-related differences among level walking, stair ascent, and ramp ascent were examined using one-way repeated-measures ANOVA (α = 0.05), with eta-squared (η2) reported as a measure of effect size and 95% confidence intervals provided for pairwise mean differences.

Results

Results are reported in three stages: exploratory participant-level split results, nested 5 × 5 cross-validation, and subject-independent LOSO and held-out-test results. The latter two regimes provide the primary evidence for cross-subject generalization, whereas the exploratory split results are reported to characterize model behavior and should not be interpreted as final generalization estimates.

Preliminary statistical feature ranking

A preliminary exploratory statistical analysis was conducted on the full set of 120 EMG-derived features, extracted from three lower-limb muscles: Tibialis Anterior (TA), Soleus (SOL), and Gastrocnemius Lateralis (GL). To avoid bias toward any single importance metric, a composite importance score was computed by combining three complementary methods: (i) Random Forest (Gini) importance, capturing non-linear feature interactions; (ii) Mutual Information, a model-free measure of non-linear dependence; and (iii) the ANOVA F-statistic, capturing linear class separability. Each method’s scores were min–max normalized to 1 and averaged with equal weights.

The resulting ranking was led by GL_LD (0.898), TA_SD (0.739), SOL_LD (0.703), and SOL_MYOP (0.626). The top-ranked features were distributed across all three muscle groups rather than concentrated in one, indicating that the discriminative signal is physiologically distributed rather than muscle-specific. The cumulative importance curve showed that 50% of total importance was concentrated within the top 29 features, and 80% within the top 67 features — evidence of substantial redundancy among lower-ranked features.

For exploratory characterization only, the top 72 features (≈83.6% of total cumulative importance) were retained as the statistically pre-filtered candidate pool for the subsequent feature-selection stage (Appendix A, Table A2). The muscle-wise distribution of these 72 features was balanced — 25 from GL, 25 from TA, and 22 from SOL — with GL contributing the largest share of cumulative importance (≈39%), followed by TA (≈34%) and SOL (≈27%).

The ranking shows that discriminative information is distributed across the tibialis anterior, soleus, and gastrocnemius lateralis rather than concentrated in one muscle. The predominance of features from multiple signal domains indicates that locomotor discrimination depends on complementary amplitude, variability, complexity, and temporal descriptors.

Comparison of features across locomotor tasks

Descriptive statistics

Appendix A, Table A3 summarizes descriptive statistics mean ± standard error of the mean (SEM) and coefficient of variation (CV %) for the top 29 features (min-max normalized to 1) across the three locomotor tasks. Feature magnitudes tended to be highest during level walking and lower during stair and ramp ascent, while variability (CV%) was generally greater during the two inclined/elevated tasks. Tibialis anterior amplitude-related features (TA_SD, TA_VO, TA_MAD, TA_LTKEO) showed the lowest CV during level walking (6.7-16.2%), reflecting a highly stereotyped activation pattern for steady gait, whereas several soleus and gastrocnemius indices (SOL_ASM, GL_ASM, SOL_DWT_db4) exceeded 60-90% CV during stair/ramp ascent, suggesting more heterogeneous, task-dependent recruitment on inclined surfaces. GL_LD showed an extreme CV during stair (199.8%) and ramp ascent (168.8%), reflecting near-sparse activation (low mean amplitude) rather than genuine measurement instability. Table A3 Descriptive statistics of normalized EMG features.

Inferential statistics: ANOVA, post-hoc tukey HSD, and effect size

A one-way repeated-measures ANOVA was performed for each of the top 29 features across the three tasks using participant-level task means; all 29 differed significantly (F = 6.38–69.09; 27 at p < 0.001, the remaining two at p < 0.01). Because the omnibus ANOVA only indicates that at least one task differs without identifying which pair(s) drive the effect, Tukey’s Honestly Significant Difference (HSD) post-hoc test was applied for all three pairwise contrasts (LW vs SA, LW vs RA, SA vs RA), with family-wise error controlled at α = 0.05 within each feature. Because 29 features were tested, these inferential results are interpreted as exploratory unless a between-feature multiplicity correction is applied. Effect size (η2) was computed for each feature as the proportion of total variance attributable to task (Cohen’s thresholds: ≈0.01 small, ≈0.06 medium, ≥ 0.14 large).The complete inferential results for the 29 highest-ranked features are reported in Appendix A, Table A4.

The most common discrimination pattern (15 of 29 features, including GL_LD, TA_SD, SOL_LD, and TA_MAD) was one in which LW differed significantly from both SA and RA, while SA and RA were not distinguishable from each other — consistent with level walking being the most electromyographically distinct task. Only three features — SOL_MYOP, GL_WL, and GL_IEMG — significantly discriminated all three pairwise comparisons, making them of particular interest for full three-class discrimination. A smaller subset (GL_ASS, GL_CARD, GL_SSC) was primarily sensitive to ramp ascent, while GL_ASM, TA_ASM, SOL_DWT_db4, and TA_MNF were primarily sensitive to stair ascent.

Every one of the 29 features exceeded the η2 = 0.14 threshold for a large effect (range 0.183–0.708; mean η2 ≈ 0.37), confirming that these differences are practically, not merely statistically, meaningful. GL_LD again ranked highest (η2 = 0.708, F = 69.09), consistent with its top rank in the composite importance score, followed by SOL_MYOP (0.558), GL_MYOP (0.541), and SOL_LD (0.500). Overall, the LW-vs-SA and LW-vs-RA contrasts reached significance for the large majority of features (24/29 and 23/29, respectively), whereas SA-vs-RA reached significance for only 10/29 — indicating that stair and ramp ascent elicit comparatively similar muscle-activation profiles relative to level walking.

The effect-size panel confirms that the statistically significant differences are also practically meaningful, while the Tukey heatmap identifies which task pairs drive the separation. The stronger level-walking contrasts relative to stair-versus-ramp contrasts indicate that ascent tasks share a partially overlapping activation strategy.

Temporal muscle activation patterns across gait phases

To further characterize lower-extremity locomotor control, the temporal activation patterns of TA, SOL, and GL were evaluated across the heel-strike, mid-stance, and toe-off phases (Table 9) 24. This analysis was performed independently of, and prior to, feature extraction for classification. The stance phase was analytically divided into three functional regions—early stance (around the conventionally defined heel-strike point), mid-stance, and late stance (ending at the conventionally defined toe-off point)—using stance-phase percentages reported in the literature 33. The current acquisition system did not record device-side packet-level timestamps or externally referenced gait-event markers, and therefore these phase boundaries were not obtained through direct heel-strike or toe-off event detection. Gait-phase segmentation was therefore used solely to characterize physiological muscle-activation behavior and was not incorporated as input to the machine-learning classifier; classification instead relied on features extracted from the complete EMG recordings of all three muscles 24,34. This separation was deliberate: using literature-derived phase percentages as classifier input would have introduced timing assumptions unverified for the present cohort, whereas extracting features from the complete, unsegmented signal allows the classifier to exploit the full temporal information present in each participant’s recording, independent of any external phase-boundary assumption.

Table 9 Mean muscle-activation amplitudes across literature-defined gait-phase regions and locomotor tasks.

Full size table

Level walking

During level walking, tibialis anterior (TA) dominated at heel strike (62.6 ± 9.1%), consistent with its established role in eccentric plantarflexion control at initial contact. During mid-stance, plantarflexor activation increased substantially (Soleus: 43.2 ± 6.2%; GL: 41.6 ± 5.8%), followed by further elevation at toe-off (Sol: 44.1 ± 4.0%; GL: 45.2 ± 8.7%) with near-absent TA activity (3.5 ± 5.1%). This continuous plantarflexor engagement throughout stance phase corresponds to the highest amplitude feature observed across tasks (RMS: 0.859; MAV: 0.808).

Stair ascent

Stair ascent was characterized by a marked temporal redistribution of TA activation, with peak activity shifting from heel strike to mid-stance (67.4 ± 3.2%). Plantarflexor activity during mid-stance was minimal (Sol: 2.0 ± 0.7%; GL: 2.3 ± 1.0%), followed by a propulsive burst at toe-off (Sol: 43.7 ± 5.5%; GL: 33.9 ± 8.9%). This phasic activation pattern, periods of near-silence punctuated by abrupt bursts — is reflected in the elevated Waveform Length observed for this task (0.742 ± 0.048).

Ramp ascent

Ramp ascent exhibited an intermediate TA activation at heel strike (38.4 ± 2.8%), with simultaneous activation of all three muscles during mid-stance (TA: 18.5 ± 2.7%; Sol: 28.8 ± 2.9%; GL: 10.1 ± 3.1%), representing a multi-muscle co-activation pattern not observed during the other two tasks. At toe-off, Sol activation reached its highest value across all tasks (50.7 ± 6.2%). The co-activation pattern during mid-stance was associated with the highest inter-subject variability across all conditions (CV range: 51.5–63.9%), consistent with the statistically confirmed heteroscedasticity.

Influence of normalization on classification performance

Algorithm-specific normalization responses

The impact of three normalization strategies—Z-score (ZSN), min–max (MMN), and %MVC-based (MVCN)—plus an unnormalized baseline (WN) was evaluated across eight classifiers (Fig. 10). Without scaling, Logistic Regression (LR) and Support Vector Machines (SVM) yielded 48.90% and 53.44% accuracy, respectively, while k-Nearest Neighbors (KNN) achieved 71.83%, confirming that raw sEMG amplitudes lack cross-subject scale invariance due to inherent physiological variation. Tree-based and ensemble architectures were markedly more resilient to unscaled inputs, with Random Forest (RF) and XGBoost both maintaining 91.70% accuracy and Decision Trees (DT) operating at 89.68%.

Fig. 10

Classification accuracy and AUC across normalization methods without feature selection.

Normalization produced substantial gains for margin- and distance-dependent classifiers: ZSN raised LR accuracy from 48.90% to 86.67% and SVM from 53.44% to 95.04%, while MMN raised KNN performance from 71.83% to 86.67% (an absolute gain of 14.84 percentage points, corresponding to a 20.7% relative improvement). Ensemble methods, by contrast, showed minimal sensitivity to the specific scaling scheme—RF shifted only slightly to 91.96%, and XGBoost remained essentially static at 91.67%—reflecting their inherent algorithmic stability across feature scales.

Statistical validation of normalization effects

A series of one-way ANOVAs confirmed statistically significant normalization effects for seven of the eight classifiers (p < 0.05); XGBoost was the sole exception (F = 2.18, p = 0.089), consistent with the rank-order invariance of tree-based node-splitting criteria. The most pronounced sensitivities to amplitude scaling were found for LR (F = 15.67, p < 0.001), SVM (F = 12.34, p = 0.001), and KNN (F = 8.92, p = 0.003). ZSN emerged as the overall superior strategy, though MMN yielded peak performance for both the neural-network architecture (92.1%) and KNN (86.70%), and MVCN delivered the best result exclusively for LR (93.80%).

Dimensionality reduction via feature selection

A total of 120 features (40 per muscle) were extracted from the three EMG channels. Three feature-selection strategies—Genetic Algorithm (GA), Principal Component Analysis (PCA), and a hybrid PCA-GA approach—were applied, each evaluated under ZSN, MMN, and MVCN normalization.

Genetic algorithm feature selection

GA-based selection substantially improved classification accuracy across all classifiers (Figs. 11, 12), with the largest gains observed for LR (48.90% → 98.33% under ZSN) and SVM (53.44% → 96.33%). Under ZSN, GA reduced the feature set from 120 to 32 features—a 73.3% dimensionality reduction—while preserving balanced muscle representation: GL (12 features, 37.5%), SOL (11, 34.4%), and TA (9, 28.1%). Ensemble methods showed more modest but consistent gains (RF: 91.67% → 94.00%; XGBoost: 91.70% → 93.39%).

Fig. 11

Performance of GA-based feature selection across normalization methods.

Fig. 12

Classification accuracy with GA-based feature selection across normalization methods.

Seven features were consistently selected across all normalization conditions—TA_SSC, TA_DAMV, and TA_DASDV (tibialis anterior); SOL_MMAV and SOL_SKEW (soleus); and GL_LD and GL_MYOP (gastrocnemius lateralis)—indicating normalization-independent discriminative utility. In contrast, frequency-domain features (log-TKEO, wavelet descriptors) exhibited normalization-dependent selection patterns.

PCA-based feature selection

PCA reduced the 120-dimensional feature space to nine principal components under each normalization scheme, as shown in Figs. 13 and 14. SVM benefited most from PCA, improving from 33.33% to 93.33% with ZSN. NNET reached peak accuracy of 93.33% under MMN. Decision Tree performance declined under PCA (89.68% to 76.67% with ZSN), indicating sensitivity to variance-based transformations that may obscure axis-aligned decision boundaries.

Fig. 13

Performance of PCA-based feature selection across normalization methods.

Fig. 14

Classification accuracy with PCA-based feature selection.

ZSN and MMN produced the most favorable PCA outcomes. RF and XGBoost showed negligible response to PCA, consistent with their insensitivity to feature correlation structure.

Hybrid PCA-GA feature selection

The hybrid PCA-GA strategy, combining PCA-derived orthogonal components with GA evolutionary optimization, produced results closely aligned with standalone PCA as illustrated in Figs. 15 and 16. Under Z-score normalization, GA selected six of nine PCA components, representing a 95.0% reduction from the original 120 features. Under MVC normalization, seven components were retained (94.2% reduction), while Min–Max normalization yielded the most compact solution (four components; 96.7% reduction). Performance improvements over the WN baseline were comparable to those of standalone PCA (LR: + 160.2%; SVM: + 180.0%; KNN: + 18.4%), whereas Decision Tree again exhibited accuracy degradation, consistent with its behavior under preceding configurations. Crucially, no classifier demonstrated meaningful gains over standalone GA-based selection, suggesting that the two-stage PCA → GA pipeline introduces computational overhead without a proportionate discriminative benefit, an outcome indicative of feature space saturation when GA alone is applied.

Fig. 15

Comparative evaluation of the hybrid PCA–GA framework versus standalone GA and PCA across normalization methods.

Fig. 16

Classification accuracy with hybrid PCA–GA feature selection across normalization methods.

Final evaluation of GA-optimized classifiers

The three classifiers selected in “Dimensionality reduction via feature selection” section were carried forward for final evaluation using the participant-level validation/test partition defined in “Participant characteristics” section. No additional participant-level data were introduced at this stage.

Impact of normalization on final model performance

The normalization strategy applied to the sEMG features had a decisive influence on classifier performance, with Z-score normalization (ZSN) consistently yielding the strongest results across all evaluated models (Fig. 17). Without normalization, LR and SVM produced near-chance outputs (validation: 48.9% and 53.4%; test: 35.4% and 46.3%, respectively), while RF retained a comparatively high unnormalized validation accuracy of 91.7%, reflecting its inherent structural robustness to feature scaling. Following ZSN normalization, all three classifiers achieved validation and test accuracies above 92%.

Fig. 17

Impact of normalization techniques on classification accuracy for GA-optimized classifiers.

LR showed the largest gain from normalization, improving by up to 59.0 percentage points over its unnormalized baseline and reaching 98.3% validation and 94.4% test accuracy under ZSN. RF improved further under ZSN, reaching 94.0% validation and 93.1% test accuracy — a modest but consistent gain over its unnormalized baseline, reflecting stable generalization across subjects. SVM was comparatively robust to the choice of scaling method, attaining 96.3% validation accuracy under ZSN and near-equivalent test accuracy under both ZSN and MMN (92.7% vs. 92.9%), supporting its reliability as a cross-subject classifier largely independent of the specific amplitude-scaling technique used.

Critically, all three GA-optimized classifiers under the prespecified ZSN participant-level split maintained validation-to-test accuracy gaps below five percentage points — evidence of cross-subject generalization without overfitting, a key criterion for the eventual clinical translation of wearable monitoring systems.

Multi-metric classification performance

Figure 18 presents the performance of the three selected classifiers — LR, SVM, and RF — across four evaluation metrics (Accuracy, Precision, Recall, and F1-Score) under Z-score normalization. LR achieved the highest overall performance across all four metrics, with SVM close behind; RF remained competitive despite its comparatively lower sensitivity to normalization noted in “Impact of normalization on final model performance” section. The near-identical values recorded across metrics for each classifier indicate well-calibrated, balanced classification behavior across the three locomotor classes, with no systematic bias toward precision at the expense of recall, or vice versa.

Fig. 18

Classification-performance comparison of GA-optimized classifiers across evaluation metrics.

Taken together, these multi-metric results confirm that the balanced accuracy reported in “Impact of normalization on final model performance” section is not an artifact of class imbalance, reinforcing ZSN combined with GA-optimized feature selection as the preferred configuration for locomotor-task classification.

While these results are encouraging, they rest on a single validation/test split. Given the modest participant cohort available, such a design is inherently susceptible to optimistic bias: both the reported accuracy and the genetic-algorithm-driven feature-selection process may partly reflect the particular participants assigned to each split, rather than performance that would generalize reliably to unseen individuals.

To assess robustness against this possibility, the feature-selection process was subjected to a nested cross-validation scheme, and the resulting feature subsets were evaluated using subject-independent leave-one-subject-out (LOSO) cross-validation — the most stringent test of generalization achievable at this cohort size, detailed in the following sections.

Nested cross-validation refinement of the GA-selected feature subset: a 44-feature solution

To this end, the GA-based feature-selection procedure described in “Nested genetic algorithm for feature selection” section was re-run under a nested cross-validation scheme. Beyond addressing the single-split bias noted above, this re-selection also served to reduce the dimensionality of the original feature ranking used in the embedded logistic-regression pipeline. The nested procedure converged on a 46-feature subset drawn from the original 120-feature space; two of these — zc_gl and zc_sol — exhibited zero variance across the entire dataset and were excluded, leaving 44 analytically valid features for downstream ranking and statistical characterization.

A composite importance score was computed for the 44-feature set using the same procedure as the original 72-feature ranking in “Preliminary statistical feature ranking” section—the equally weighted average of min–max normalized Random Forest (Gini) importance, Mutual Information, and the ANOVA F-statistic. Appendix A, Table A5 reports the full 44-feature ranking together with each component score. The complete 44-feature ranking is reported in Appendix A, Table A5. All subject-independent performance claims were based on feature selection and hyperparameter tuning performed within the corresponding training folds.

As shown in Fig. 19, the ranking is led by fzc_gl (0.810), myop_gl (0.743), vare_ta (0.679), and var_ta (0.632) — again spanning all three muscles, echoing the physiologically distributed discriminative signal identified in “Preliminary statistical feature ranking” section. GL-derived features contribute disproportionately to the upper ranks, while the lowest-ranked features (mmav_gl, kurt_ta) fall below a composite score of 0.05, foreshadowing the redundancy addressed by the correlation-based pruning step.

Fig. 19

Nested-GA feature characterization: (a) composite importance scores for the 44 valid features and (b) feature-level task-response heatmap.

Descriptive statistics (mean ± SEM, CV%) and one-way ANOVA with Tukey HSD post-hoc comparisons were computed for all 44 features across the three locomotor tasks, following the same procedure applied to the top-29 set in “Comparison of features across locomotor tasks” section. Thirty-eight of the 44 features (86%) exceeded the large-effect threshold (η2 ≥ 0.14); the remaining six — sd_sol, mad_gl, iemg_ta, rms_gl, card_sol, and kurt_ta — showed small-to-negligible effects, anticipating the pruning applied in “Redundancy elimination via pairwise correlation pruning” section. The complete pairwise post-hoc significance pattern is visualized in Fig. 19; the leading feature, fzc_gl, again shows the strongest overall discrimination (F = 60.80, p < 0.001, η2 = 0.681), consistent with its top composite-importance rank.

Under the nested cross-validation procedure, three validation regimes were compared for this 44-feature set: Genetic-Optimized K-Fold, Nested 5 × 5 CV, and Leave-One-Subject-Out (LOSO). Under LOSO — the strictest, fully subject-independent regime — accuracy ranged from 81.1% (NNET) to 89.0% (RF), with RF, LR, and SVM the strongest performers (87.2–89.0%). Per-subject LOSO accuracy breakdowns across all 20 training/cross-validation participants (S1–S20) are provided in Fig. 20.

Fig. 20

Radar-chart comparison of classification accuracy for the 44-feature nested-GA representation across Genetic-Optimized K-Fold, Nested 5 × 5 cross-validation, and LOSO validation.

Figure 20 demonstrates that the nested-GA feature set remains competitive as the evaluation becomes progressively more stringent. The close relationship among genetic-optimized K-fold validation, nested 5 × 5 cross-validation, and LOSO performance indicates that the selected representation retains discriminative information beyond a single random split, while the lower LOSO values quantify the expected generalization penalty when participant-specific information is removed. This pattern supports the use of subject-independent validation as the more credible estimate of deployment performance.

Redundancy elimination via pairwise correlation pruning

Building on the 44-feature nested-GA set, pairwise Pearson correlation analysis with a prespecified threshold of |ρ|> 0.90 identified and removed 27 redundant features, yielding a final set of 17 high-importance features with low pairwise redundancy. One-way ANOVA with Tukey HSD post-hoc comparisons confirmed strong task discrimination for the retained representation: fzc_gl (F = 60.80, p < 0.001, η2 = 0.681), ld_sol (F = 28.65, p < 0.001, η2 = 0.501), and lcov_gl (F = 25.48, p < 0.001, η2 = 0.472) showed large effects, while 15 of the 17 retained features remained statistically significant discriminators (p < 0.05). The resulting 17-feature set was therefore retained as a lightweight, resource-constrained alternative rather than being presented as a model that formally meets the reviewer’s suggested ≥ 95% LOSO threshold.

The complete 17-feature descriptive table is reported in Appendix A, Table A6. The corresponding composite importance ranking is provided in Appendix A, Table A7. The complete inferential statistics, effect sizes, and Tukey HSD comparisons for the 17 retained features are reported in Appendix A, Table A10.

Table 10 Mathematical Definitions of the Extracted EMG Features.

Full size table

Thirteen of the 17 retained features (76%) preserved a large effect size (η2 ≥ 0.14) after pruning — a modest, expected decrease from 86% (38/44) before pruning, consistent with removing correlated rather than uninformative features. Fifteen of the 17 retained features remained statistically significant task discriminators (p < 0.05), whereas sd_sol and card_sol were retained because of their non-trivial composite importance and low correlation with the remaining features; they are therefore considered candidates for further conservative pruning, and fzc_gl, ld_sol, and lcov_gl retained the largest effect sizes (η2 = 0.681, 0.501, and 0.472, respectively). The complete pairwise significance pattern is shown in Fig. 21.

Fig. 21

Correlation-pruned feature characterization: (a) composite importance scores for the 17 retained features and (b) feature-level task-response heatmap.

Comparative classification performance of the 44-feature and 17-feature subsets across validation regimes

In the post-selection sensitivity analysis, the 17-feature representation matched or exceeded the 44-feature representation for all eight classifiers under the reported LOSO summaries; these values should not be interpreted as a fully nested estimate of pruning performance unless the pruning step is repeated within every outer training fold; performance differences were mixed under the less stringent Genetic-Optimized K-Fold and Nested 5 × 5 regimes. This was achieved despite a 61% reduction in dimensionality (e.g., DT: 88.52% vs 84.98%; SVM: 92.40% vs 87.24%; XGBoost: 86.61% vs 83.74%). The 17-feature DT classifier reached 95.00% accuracy under Nested 5 × 5 CV. The complete cross-regime comparison is provided in Appendix A, Table A8. Figures 22a and b show the corresponding subject-level LOSO accuracy patterns for the 44-feature and 17-feature representations, respectively.

Fig. 22

Per-subject LOSO accuracy across eight classifiers (a) 44-feature nested-GA representation (S1–S20) and (b) 17-feature nested-GA representation (S1–S20)

The 44-feature heatmap shows generally high but heterogeneous subject-level performance. Random Forest, SVM, and Logistic Regression maintain the strongest overall accuracy profiles, whereas NNET and Naive Bayes show greater subject-to-subject variability.

Compared with Fig. 22a, the 17-feature representation preserves or improves the accuracy profile for most subjects and classifiers while using substantially fewer descriptors. The strongest and most consistent performance is concentrated in the Logistic Regression, Random Forest, and SVM columns, supporting the conclusion that correlation pruning reduced dimensionality without sacrificing subject-independent discrimination.

Benchmarking handcrafted features against end-to-end deep learning: 1D-CNN and BiLSTM baselines

Two lightweight deep-learning architectures (1D-CNN and BiLSTM) were trained directly on raw, temporally aligned three-channel sEMG windows — without handcrafted features — and evaluated under the same subject-independent partitioning used for the traditional ML pipeline: a held-out 5-participant test set, 5 × 5 nested cross-validation, and LOSO over the 20 training/cross-validation participants.

Under LOSO, the 1D-CNN reached 91.39% ± 12.92% accuracy and the BiLSTM reached 87.72% ± 16.16% — closely matching the strongest traditional classifier overall (17-feature SVM: 92.40%) and exceeding the best 44-feature LOSO result (RF: 89.02%) — but with substantially larger between-subject variance (13–16 percentage points, versus single digits for the traditional pipeline), reflecting the smaller effective per-fold training sample available to the deep networks. The deep-learning baseline comparison is reported in Appendix A, Table A9.

On the independent 5-participant test set (S21–S25), mean per-subject accuracy was 85.72% ± 15.77% (1D-CNN) and 79.56% ± 16.02% (BiLSTM), confirming that the generalization gap relative to the feature-based pipeline persists under fully held-out evaluation. In terms of model complexity, the 1D-CNN has 154,307 trainable parameters and requires ≈2.50 million FLOPs per inference; the BiLSTM has fewer parameters (77,187) but is substantially more computationally expensive per inference (≈9.08 million FLOPs), reflecting the sequential nature of the recurrent architecture. The substantially smaller input and feature-extraction burden of the traditional 17-feature pipeline provides an accuracy-per-FLOP advantage relevant to embedded and wearable deployment, even though the deep-learning baselines achieve comparable mean LOSO accuracy.

Discussion

Synthesis of principal findings

This study evaluated the feasibility of classifying three locomotor tasks — level walking, stair ascent, and ramp ascent — from surface electromyographic (sEMG) signals recorded from the tibialis anterior, soleus, and gastrocnemius lateralis using a systematic, multi-stage machine-learning pipeline. Four principal findings emerged. First, the discriminative information carried by the 120 extracted EMG features is physiologically distributed across all three muscles rather than concentrated in a single muscle, with gastrocnemius lateralis contributing the largest share of cumulative importance (≈39%). Second, all 29 most important features differed significantly across the three tasks with large effect sizes (η2 ≥ 0.14), and level walking proved to be the most electromyographically distinct task, whereas stair and ramp ascent elicited comparatively similar activation profiles. Third, the choice of normalization strategy was decisive for scale-sensitive classifiers — Z-score normalization raised logistic regression accuracy from 48.90% to 98.33% under genetic-algorithm feature selection — while tree-based ensembles proved remarkably resilient to unscaled inputs. Fourth, genetic-algorithm feature selection substantially outperformed PCA and the hybrid PCA-GA pipeline, and a subsequent correlation-based pruning step distilled the feature set to 17 features that matched or exceeded the performance of the full 44-feature nested-CV set under leave-one-subject-out evaluation (92.40% vs 87.24% for SVM), with all normalized models maintaining validation-to-test accuracy gaps below five percentage points.

Physiological interpretation of the task-related EMG differences

The exploratory statistical separation among tasks documented in “Comparison of features across locomotor tasks” section is consistent with the known biomechanics of level gait versus inclined/elevated ascent. During level walking, tibialis anterior activation peaked at heel strike (62.6 ± 9.1% MVC), reflecting its classical role in controlling the rate of ankle dorsiflexion at initial contact, while the plantarflexors (soleus and gastrocnemius) were engaged continuously throughout mid-stance and toe-off, consistent with their function in body-weight support and forward propulsion 1. The elevated amplitude-related features (e.g., TA_RMS, TA_LTKEO) and their low cross-trial variability during level walking (CV 6.7–16.2%) reflect this highly stereotyped activation pattern.

During stair ascent, tibialis anterior activation shifted dramatically from heel strike to mid-stance (67.4 ± 3.2% MVC), a redistribution that has been attributed to the greater dorsiflexion moment required to raise the body over the step during single-limb support, combined with near-silent plantarflexor activity at mid-stance followed by a propulsive burst at toe-off 2. This phasic, burst-like pattern explains both the elevated waveform length and the distinct feature profile of the stair condition. Ramp ascent produced a genuinely different control strategy: a multi-muscle co-activation of all three muscles at mid-stance — not observed in either of the other two tasks — and the highest soleus activation at toe-off across all conditions (50.7 ± 6.2% MVC). This co-activation is plausibly related to the sustained anterior–posterior stability demands of incline gait, during which the trunk must be controlled against gravity while the leading limb accepts weight, and it coincides with the highest inter-subject variability in the dataset, suggesting that subjects adopted individualized stabilization strategies on the ramp 3.

The finding that only three features (SOL_MYOP, GL_WL, GL_IEMG) discriminated all three pairwise task contrasts is physiologically informative: the waveform-length and myopower descriptors capture both the amplitude and temporal texture of the signal, which are precisely the dimensions most affected by the phase shifts and co-activation patterns described above. The near-sparse activation of GL_LD during stair and ramp ascent (with extreme CV values of 199.8% and 168.8%) likewise illustrates that features carried by a muscle that is largely inactive in two of the three tasks derive their discriminative power from task absences rather than task magnitudes — an effect that classifiers can exploit but that underscores the importance of the CV statistics reported in Appendix A, Table A3.

Why normalization matters — and why ensembles are robust

The normalization analysis (Sect. “Influence of Normalization on Classification Performance”; Fig. 10) quantifies a well-known but often under-reported property of sEMG-based classification: raw amplitude features lack cross-subject scale invariance, so margin- and distance-dependent classifiers (LR, SVM, KNN) perform near chance without scaling (48.90–71.83%), whereas Z-score normalization recovers 86.67–95.04%. This is consistent with the large inter-subject differences in electrode placement, subcutaneous tissue thickness, and muscle size that dominate raw sEMG amplitudes 4. As shown in Fig. 18, the validation-to-test gap for Z-Score-normalized models under this initial Genetic-Optimized K-Fold split was already narrow (e.g., LR: 98.3% validation vs. 94.4% test; SVM: 96.3% vs. 92.7%; RF: 94.0% vs. 93.1%), offering an early indication of stable generalization even before the more stringent nested and subject-independent evaluations described in “Feature selection: genetic algorithms add value, pca adds complexity” section.

Two results deserve emphasis. First, Z-score normalization was superior overall, even for logistic regression, for which %MVC normalization delivered the single best result (93.80%); the practical implication is that when %MVC reference contractions cannot be reliably elicited — a common limitation in unsupervised or clinical populations — ZSN remains a safe default. Second, the one-way ANOVAs confirmed that normalization has no statistically detectable effect on XGBoost (F = 2.18, p = 0.089) and minimal effect on RF, corroborating the theoretical insensitivity of rank-order-based tree splitting to monotonic scaling 5. For wearable deployments, this ensemble robustness is valuable: it provides a graceful fallback when preprocessing pipelines fail or when scaling parameters estimated from training data are mismatched at inference time.

Feature selection: genetic algorithms add value, PCA adds complexity

The comparison of the three dimensionality-reduction strategies reveals a clear performance hierarchy. The genetic algorithm (GA) delivered the largest and most consistent gains, raising logistic regression accuracy from 48.90% to 98.33% (Fig. 17) and SVM to 96.33%, while simultaneously reducing the feature space by 73.3% to a balanced 32-feature set. Seven features — TA_SSC, TA_DAMV, TA_DASDV, SOL_MMAV, SOL_SKEW, GL_LD, and GL_MYOP — were selected consistently across all normalization conditions, constituting a normalization-invariant core that spans all three muscles and integrates amplitude- (MMAV), shape- (SSC, skewness), and nonlinear- (log-TKEO-domain) descriptors. This composition is consistent with the multi-metric importance structure identified in “Preliminary statistical feature ranking” section, reinforcing that no single feature domain alone captures the full discriminative signal.

PCA, by contrast, offered no advantage over GA for any classifier and actively degraded decision-tree performance. This outcome is attributable to the fundamental nature of principal components as variance-optimal linear projections that do not respect class boundaries: axis-aligned decision surfaces can be destroyed even when discriminative variance is nominally preserved 6. The hybrid PCA–GA pipeline inherited these same limitations — no classifier gained meaningfully over standalone GA despite an additional dimensionality reduction of up to 96.7% — indicating that the two-stage approach adds computational overhead without a proportionate discriminative benefit. For resource-constrained wearable inference, the standalone GA solution therefore represents the optimal accuracy-per-computation trade-off.

The subsequent refinement stages (Sects. “Nested Cross-Validation refinement of the GA-selected feature subset: A 44-feature solution”–”Redundancy elimination via pairwise correlation pruning”) provide a consistency check under subject-independent evaluation, and reveal a consistent, expected decline in accuracy as evaluation stringency increases and the risk of subject-level leakage is progressively reduced. Re-running the GA under nested cross-validation — necessary to correct the original single-split validation and prevent the feature-selection search from being informed by held-out data — produced a 44-feature set whose top-ranked features (fzc_gl, myop_gl, ld_sol) echoed the distributed, GL-weighted importance structure identified in “Preliminary statistical feature ranking” section, albeit at lower accuracy than the original 32-feature K-Fold result (Fig. 17), reflecting the removal of this optimistic bias.

A cross-method comparison of the feature sets selected across the three independent selection strategies — the composite statistical ranking (Sect. “Preliminary statistical feature ranking”, top 29 of 120), the single-split GA (Section “Genetic algorithm feature selection”, 32 of 120), and the nested-CV GA (Sect. “Nested cross-validation refinement of the GA-selected feature subset: A 44-feature solution.”, 44 of 120) — provides an additional consistency check on the discriminative feature space. Because these three procedures differ fundamentally in their optimization objective (multi-metric statistical importance vs. classifier-embedded evolutionary search vs. subject-independent evolutionary search), substantial disagreement between the resulting sets would be expected a priori. Nonetheless, five features — SOL_LD, GL_LCOV, GL_MYOP, SOL_MMAV, and GL_ASM — were retained by both the statistical top-29 ranking and the nested-CV GA 44-feature set, and all five survived the subsequent correlation-based pruning into the final 17-feature set. All five features originate from the plantarflexor group (soleus and gastrocnemius), reinforcing the physiological interpretation in “Physiological interpretation of the task-related EMG differences” section that plantarflexor descriptors carry the most consistently transferable discriminative signal across both amplitude-domain and evolutionary selection criteria. Conversely, the fact that only 10 of the 29 statistically top-ranked features (≈34%) reappeared in the nested-CV GA set underscores that composite statistical importance and embedded, classifier-driven selection are only partially overlapping proxies for discriminative utility — a distinction with direct methodological implications for future feature-selection pipelines that rely on a single ranking criterion in isolation (Fig. 23).

Fig. 23

Deep-learning LOSO baselines across the 20 validation participants: (a) 1D-CNN and (b) BiLSTM per-subject accuracy, precision, recall, and F1-score.

Correlation-based pruning of this 44-feature set to 17 features subsequently recovered accuracy under the same K-Fold protocol by eliminating redundant (|ρ|> 0.90) feature pairs rather than informative ones, with 76% of the pruned set retaining large effect sizes.

Applying progressively stricter subject-independent protocols to this 17-feature set — Nested 5 × 5 CV, followed by LOSO, followed by the fully independent held-out test set — produced a further, stepwise decline in accuracy at each stage, consistent with the increasing elimination of subject-level information leakage. This convergence across evaluation regimes is summarized in Fig. 24, which presents accuracy by classifier across the genetic-optimized K-Fold, nested 5 × 5 CV, and LOSO protocols. The consistent ordering observed across nearly all classifiers (K-Fold ≥ Nested 5 × 5 ≥ LOSO) confirms that this tightening reflects a systematic property of the evaluation design rather than an unstable or classifier-specific artifact.

Fig. 24

Radar-chart comparison of classification accuracy for the 17-feature correlation-pruned representation across the three validation regimes.

This radar-chart summary provides a compact visual comparison of classifier accuracy across the three validation regimes and complements the 44-feature summary presented earlier.

The monotonic decrease from genetic-optimized K-fold validation to nested cross-validation and LOSO reflects the increasing protection against subject-level information leakage. The consistency of this ordering across classifiers strengthens the interpretation that the performance reduction is a property of evaluation stringency rather than random model instability.

The transparent reporting that neither feature set reached the 95% LOSO threshold (best: 17-feature SVM at 92.40%) is a deliberate methodological choice: it positions the 17-feature model as a lightweight, honestly validated alternative rather than an over-claimed universal solution.

Generalization, overfitting control, and the deep-learning baselines

For the three GA-optimized classifiers under the prespecified ZSN participant-level split, validation-to-test accuracy gaps remained below five percentage points and converged with the ROC analysis (Fig. 25) to are consistent with limited overfitting, although they do not prove the absence of overfitting given the small training cohort (20 participants). AUC values exceeded 0.98 for all classifiers in both validation and testing phases, with LR achieving the highest discrimination for level walking (0.998), SVM performing marginally best for stair ascent (0.997), and RF showing the most balanced performance across all three classes (AUC range: 0.985–0.992); critically, the inter-phase AUC deviation remained below 2% across all models, confirming that this stability holds at the class level and not merely in aggregate accuracy.

Fig. 25

ROC curves for LR, SVM, and RF during validation and testing, by locomotor class.

The ROC curves show that the classifiers preserve high class-level discrimination in both validation and testing phases. The close overlap between phases indicates that the strong aggregate accuracy is not explained solely by one favorable operating point or by a single locomotor class.

The deep-learning baselines sharpen this picture: the 1D-CNN (91.39% ± 12.92%) and BiLSTM (87.72% ± 16.16%) matched or slightly exceeded the traditional pipeline under LOSO in the mean but with two- to threefold larger between-subject variance — a contrast visible in their per-subject test-set breakdowns (Figs. 26–27), where a minority of participants (S21, S25) fall to 55–69% while the majority exceed 94% — and their overall test-set performance (85.72% and 79.56%) lagged the feature-based pipeline more clearly.

Fig. 26

1D-CNN per-subject test-set accuracy, precision, recall, and F1-score for S21–S25.

Fig. 27

BiLSTM per-subject test-set accuracy, precision, recall, and F1-score for S21–S25.

The per-subject results reveal that the 1D-CNN achieves strong performance for most held-out participants but exhibits greater dispersion for a minority of subjects. This variability is important because it shows why mean accuracy should be interpreted together with subject-level stability in wearable applications.

The BiLSTM results show a similar pattern, with a larger spread across participants than the compact feature-based models. The lower-performing subjects suggest sensitivity to limited subject-independent training data and motivate larger multi-center cohorts for end-to-end deep learning.

This pattern is consistent with the smaller effective per-fold training sample available to end-to-end networks when handcrafted features are omitted, and with prior comparisons showing that feature-based pipelines remain competitive with deep learning on small-sample EMG corpora 7. From an engineering standpoint, the 17-feature SVM (92.40% LOSO) combines lower per-inference cost (a linear model over 17 features) with a substantially lower model complexity than either deep baseline. Direct comparisons of between-subject variance should be interpreted cautiously because the reported summaries were generated under different model families and evaluation outputs (the per-subject 17-feature LOSO analysis, with no severe subject-level failures, in contrast to the 44-feature analysis and either deep-learning baseline) and by its per-subject test-set accuracy on the same independent S21–S25 cohort used for the deep-learning baselines (Fig. 28: all eight traditional classifiers remain within 77.6–93.6%, with no subject falling to the 55–69% range seen for the 1D-CNN and BiLSTM in Figs. 26–27) — making it a favorable choice for battery-constrained wearable hardware.

Fig. 28

Independent test-set per-subject accuracy for the 17-feature representation and eight classifiers (S21–S25).

The complete accuracy comparison is reported in Appendix A, Table A8. All values are Accuracy (%). The best LOSO result overall was achieved by SVM with the 17-feature set (92.40%), reported here transparently rather than overstated, consistent with best practices in performance reporting.

Limitations and future directions

The following limitations are reported transparently rather than omitted, together with concrete plans to address them in future hardware and analysis iterations.

  • Hardware bench characterization. The noise floor and BLE packet-loss characteristics of the proposed system are currently derived from component-level design specifications rather than empirical bench testing. Noise floor has not yet been measured under short-circuited or resistive-input conditions, and packet loss has not been quantified across transmission distances. Empirical characterization, including short-input RMS-noise testing after bandpass filtering and sequence-numbered packet-loss quantification at short, intermediate, and extended range, is a priority validation step for the next hardware iteration.

  • Gait-event synchronization. The current acquisition protocol does not embed explicit device-side packet-level timestamps or externally referenced gait-event markers, such as footswitch- or IMU-derived heel-strike and toe-off signals, within the transmitted EMG data stream. Consequently, the current gait-phase analysis is descriptive and based on literature-reported stance-phase percentages rather than timestamped event detection. Incorporating synchronized event markers is identified as a priority hardware and firmware enhancement for future versions of the system.

  • Sex-inclusive validation. The cohort was restricted to healthy male participants to control for well-documented sex-related differences in lower-limb muscle morphology, subcutaneous adipose tissue thickness, and gait biomechanics that could confound sEMG amplitude normalization. Extending validation to a sex-balanced or sex-stratified cohort is identified as a priority direction for future work.

  • Cohort size and single-center recruitment. With 20 participants allocated to training/cross-validation and 5 to the independent test set, the overall cohort remains modest relative to the data requirements typically associated with deep-learning approaches (Sect. “Comparative classification performance of the 44-feature and 17-feature subsets across validation regimes”). In addition, all participants were recruited from a single institution (Suez University), which may limit the generalizability of the findings to broader populations. Multi-center recruitment with an expanded cohort is identified as a priority for future validation studies.

Conclusion

This study developed and evaluated an integrated, low-cost wireless sEMG framework that combines custom wearable hardware, systematic normalization, nested feature selection, correlation-based redundancy reduction, and Bayesian-optimized machine learning for classifying level walking, stair ascent, and ramp ascent. The principal result was that a compact 17-feature representation achieved the strongest subject-independent performance, reaching 92.40% LOSO accuracy with SVM and matching or exceeding the larger 44-feature solution across the evaluated classifiers. The analysis also showed that z-score normalization was particularly important for scale-sensitive models, whereas tree-based ensembles were comparatively robust to amplitude scaling. Physiological analyses confirmed that task-related differences were distributed across the three recorded muscles, with level walking most clearly separated from the two ascent tasks and ramp ascent characterized by mid-stance co-activation and the highest soleus activity at toe-off. The feature-based pipeline remained competitive with the 1D-CNN and BiLSTM baselines while requiring a substantially smaller computational representation. Collectively, these findings support the feasibility of an interpretable and embedded-deployable system for wearable gait assessment and rehabilitation monitoring. Translation to broader clinical use will require prospective validation in larger, sex-inclusive, multi-center cohorts, together with empirical characterization of noise, packet loss, and gait-event synchronization.

来源:Nature 健康护理主题 · nature.com