Introduction

In clinical settings, precise movement assessment facilitates the early diagnosis, treatment planning, and monitoring of neuromuscular impairments. Consequently, movement analysis has become increasingly crucial, particularly in rehabilitation medicine, as it provides a comprehensive understanding of motor function, optimizes performance, and informs individualized rehabilitation strategies1. Over the past few decades, the development of movement analysis has accelerated, particularly with the advent of muscle synergy analysis (MSA), which employs unsupervised machine learning (ML) techniques to elucidate the neurophysiological control of movement2. A key principle of MSA is to uncover the spatial and temporal coordination of the activation of multiple muscles to produce efficient movement, rather than controlling each muscle independently3. Currently, in the fields of neurological4 and musculoskeletal5 diseases, MSA has provided valuable insights into motor control mechanisms by factorizing coordinated activation into spatial and temporal features.

Among various neuromuscular impairments, stroke has been the most frequently investigated using MSA to identify its gait features, due to its high incidence rate6. Notably, stroke is the second leading cause of death worldwide7 and causes persistent post-survival motor impairments that significantly affect the quality of life8. Many studies have elucidated specific patterns of gait dysfunction in stroke patients, such as the phenomenon of synergy merging, in which multiple muscle synergies combine into a single, less efficient, and specific synergy during gait9,10. Recovery from these pathological gait patterns is therefore considered a critical target in stroke rehabilitation to restore more normalized gait.

Several prior studies have applied muscle synergy features, particularly temporal coefficients, in combination with tree-based ML models to assess motor impairment and predict rehabilitation outcomes in stroke patients. For example, Yang et al. demonstrated the utility of temporal muscle synergy features by applying random forest classifiers to characterize impairment severity during sit-to-stand tasks11 and to estimate the effects of short-term rehabilitation12. An et al. further extended this approach to kinetic assessments using handrail force data in severely affected patients13. While these studies focused on temporal features to assess functional states, our study integrates both gait-specific temporal and spatial features, placing particular emphasis on model-level interpretability for pre-rehabilitation decision-making.

By utilizing MSA along with knowledge to understand factorized features, clinicians and practitioners can develop targeted rehabilitation protocols, optimize training strategies, and refine therapeutic interventions to improve gait performance in stroke patients. However, its clinical adoption remains limited due to ambiguities in interpreting and prioritizing synergy features, which often requires substantial practitioner expertise and introduces variability in application14,15,16. Therefore, in this study, we propose that an ensemble ML framework that highlights clinically relevant synergies may facilitate effective translation into practice. To the best of our knowledge, despite the introduction of various ML models and advancements in computer-aided diagnostic systems designed to reduce subjectivity and improve the consistency in assessments using electromyography (EMG)17,18, the application of MSA for therapeutic decision-making within an ML framework remains extremely rare. In particular, our approach introduces a bidirectional decomposition strategy anchored to normative synergy templates and incorporates SHAP (SHapley Additive exPlanations)-based individualization to enhance interpretability and clinical relevance.

In addition to their use in clinical diagnostics, EMG signals have been widely applied in robotics to train ML algorithms for controlling prosthetic arms and legs. This process typically involves a series of preprocessing steps such as signal filtering, normalization, and feature extraction in both the time and frequency domains. Among these steps, segmentation window size is critical, as it influences both temporal resolution and feature reliability. Shorter windows enhance temporal precision but increase variability, while longer windows stabilize signals but may miss transient events19. Commonly extracted features include root mean square, mean absolute value, median frequency, and wavelet coefficients. These features are then used as inputs for various ML models, including support vector machines, random forests, and deep learning architectures, to predict motor intentions and classify movement patterns. However, these studies have primarily focused on improving the model performance, often overlooking the interpretability of the contributing features or the physiological relevance of the learned representations20,21.

To address these limitations, this study proposed a framework aimed at enhancing the clinical interpretability of EMG data by extracting physiologically meaningful features using MSA. These features were subsequently used to train an ensemble ML model. Feature importance metrics and SHAP analysis were applied to identify synergy components that most influenced the model’s output, enabling individualized prioritization of muscle synergies, which refers to selecting the most relevant features for each patient based on SHAP-derived importance values. Furthermore, a voting-based classification approach was adopted by integrating four distinct tree-based ML models to ensure performance robustness across varying data characteristics. To support reliable interpretation, model-specific weights were assigned based on each model’s predictive performance when aggregating SHAP values and feature importance.

Notably, the spatial and temporal features derived from MSA exhibit distinct structural and statistical characteristics. Separate voting classification models were developed for each feature type and subsequently integrated into a unified ensemble model (Fig. 1). This approach provides a unified framework for decision making by consolidating predictions from both feature sets into a single model.

Accordingly, the proposed framework may offer several stage-specific advantages along the analytical pipeline. First, the use of non-negative matrix factorization (NMF) to extract muscle co-activation patterns (spatial features) and activation timings (temporal features) allows clinicians to identify which muscle groups and gait phases may be most influential in each patient. Second, these quantified features can be translated into a preliminary prioritization scheme, guiding therapists toward muscle synergies that the model indicates as clinically significant. Third, because the framework summarizes its outputs using familiar visualization tools such as feature importance scores and SHAP plots, it supports clinical interpretation without requiring expertise in ML, potentially improving usability in practice.

Fig. 1
figure 1

Overview of the muscle synergy-driven ML framework for stroke gait rehabilitation: from EMG to SHAP-informed therapeutic decision-making.

Full size image

Results

Extracted muscle synergy

Four muscle synergies were extracted from all participants based on the criterion that collectively accounted for over 90% of the EMG variance in healthy controls (variance accounted for, VAF = 0.95 ± 0.03). This number aligns with previously reported synergy structures identified during human locomotion22,23.

Numerous studies on MSA have demonstrated that NMF serves as a robust method for uncovering modular motor control strategies, as it enforces physiological non-negativity constraints and aligns closely with the structure of real-world muscle activation patterns22,23. By grouping co-activated muscles into low-dimensional modules, NMF captures the functional coordination among individual muscles, thereby enhancing the interpretability of complex motor behaviors. In this study, each extracted synergy was explicitly linked to its physiological role based on dominant muscle contributions and their established biomechanical functions during gait.

As summarized in Table1, Synergy 1 (S1) predominantly involves the hamstring muscles, including the biceps femoris (BFM) and semimembranosus (SEM), functioning primarily for deceleration and stabilization of the leg during the terminal swing and early stance phases. Synergy 2 (S2), characterized by strong activation of Rectus Femoris (RFM) and Adductor Magnus (ADD), contributes to knee extension and hip adduction, playing a key role in load acceptance and medial-lateral control during the early stance phase. Synergy 3 (S3), predominantly involving the Gastrocnemius (GCM), is critical for forward propulsion and push-off during terminal stance and pre-swing phases. Finally, Synergy 4 (S4), involving Gluteus Medius (GLU), Vastus Medialis (VMM), and Tibialis Anterior (TIB), supports pelvic stability and stance phase control, particularly during mid-stance24,25..

As shown in Fig. 2, visual comparison revealed that in patients with stroke, the muscles contributing to individual synergies (spatial features) were generally more strongly activated than in normal controls, and the timing of synergy activation (temporal features) appeared ambiguous and delayed.

Moreover, to assess the impact of stroke on synergy organization, we applied a bidirectional decomposition strategy. Extracting the spatial synergy features of stroke patients with fixed temporal features from healthy controls revealed a significant decrease in VAF (0.19 ± 0.16), indicating severely disrupted muscle coordination post-stroke. Conversely, extracting temporal synergy features with fixed spatial patterns resulted in a relatively higher VAF (0.70 ± 0.10). These findings support the hypothesis that spatial features (muscle coordination) are more substantially affected by stroke compared to temporal features (activation timing), consistent with previous clinical observations of altered neuromuscular recruitment patterns in post-stroke gait25,26.

By clearly relating NMF-derived synergy structures to established biomechanical functions and clinical phenomena, our proposed framework significantly enhances interpretability and facilitates clinical decision-making for targeted and individualized gait rehabilitation.

More at link.