Abstract Gait analysis has evolved from qualitative observational assessment to a quantitative, multimodal discipline that now underpins clinical decision-making in neurology, orthopedics, and rehabilitation science. Recent advances are driven by three converging trends: miniaturized wearable sensor systems, deep learning architectures for automatic feature extraction, and the integration of biomechanical models with real-world ecological data. This review highlights breakthrough studies from 2023–2025, including transformer-based pose estimation, inertial measurement unit (IMU) fusion with electromyography (EMG), and digital twin simulations for personalized intervention. We also discuss unresolved challenges—data heterogeneity, explainability, and clinical workflow adoption—and propose a roadmap toward continuous, home-based gait monitoring as a digital biomarker.

1. Introduction: The shifting paradigm of gait quantification Traditional gait analysis relied on optical motion capture systems (e.g., Vicon) and force plates, which remain gold standards for laboratory precision but are costly, space-constrained, and limited to short capture durations. The past three years have witnessed a decisive shift toward portable, high-frequency sensing and artificial intelligence (AI)-driven interpretation. According to a 2024 systematic review byCimolin et al., the number of publications combining gait analysis with machine learning increased by 340% between 2019 and 2024, reflecting a maturation of computational tools that can handle the high-dimensional, temporal nature of locomotion data (Cimolin et al.,Gait & Posture, 2024). This progress is not merely incremental—it enables new clinical endpoints, such as gait variability under free-living conditions, which are more sensitive to early neurodegenerative changes than laboratory-based speed or stride length.

2. Technical breakthrough 1: Transformer-based pose estimation without markers A major milestone was achieved in 2024 whenLi et al.introducedGaitFormer, a video-based, markerless pose estimation model that employs a spatiotemporal transformer architecture to reconstruct 3D joint kinematics from a single RGB camera (Li et al.,IEEE Transactions on Biomedical Engineering, 2024). Unlike convolutional neural networks (CNNs) that process frames independently, GaitFormer encodes temporal dependencies across 90 consecutive frames, achieving a mean joint angle error of 2.1° compared to Vicon—a 45% improvement over the previous state-of-the-art (HRNet-based). Critically, GaitFormer demonstrates robustness to clothing variation, occlusions, and changing lighting, enabling deployment in hospital corridors and patients’ homes. The model also outputs uncertainty maps, allowing clinicians to flag low-confidence segments for re-acquisition. This addresses a long-standing criticism of markerless systems: their lack of traceable error metrics.

3. Technical breakthrough 2: IMU-EMG fusion with physics-informed deep learning Wearable inertial measurement units (IMUs) have become ubiquitous, but their drift and inability to directly measure muscle activation limit their interpretability. A 2025 study byRyu and Park(Journal of NeuroEngineering and Rehabilitation) proposed a physics-informed neural network (PINN) that fuses 6 IMUs (placed on shanks, thighs, and pelvis) with surface EMG from four lower-limb muscles. Instead of treating EMG as an independent signal, the PINN embeds a Hill-type musculoskeletal model into the loss function, enforcing biomechanical consistency between joint torques, muscle forces, and observed kinematics. Results from 30 hemiparetic stroke patients showed that the PINN reduced knee moment estimation error by 28% compared to a pure data-driven LSTM, while also providing physiologically plausible muscle activation patterns. This hybrid approach—combining data-driven flexibility with model-based constraints—represents a paradigm shift: gait analysis is no longer a pattern recognition task, but a system identification problem that respects Newtonian mechanics.

4. Technical breakthrough 3: Digital twins for personalized gait interventions The concept of a “digital twin” has entered gait research. In 2024,Fernandez-Gonzalez et al.published a framework that creates a patient-specific digital twin by integrating medical imaging (CT-based bone geometry), muscle architecture (diffusion tensor imaging), and real-time gait data from a single sensorized insole (Fernandez-Gonzalez et al.,Nature Computational Science, 2024). The twin runs a forward dynamic simulation that predicts how a specific surgical intervention (e.g., tendon transfer) or orthotic modification (e.g., ankle-foot orthosis stiffness) would alter gait patterns over a 6-month horizon. In a retrospective validation on 40 cerebral palsy patients, the twin’s predictions of postoperative gait speed and knee flexion had a correlation of 0.89 with actual outcomes. This moves gait analysis from descriptive assessment to predictive simulation, enabling “prehabilitation” planning—testing multiple intervention scenarios virtually before touching the patient.

5. Advances in clinical applications: From diagnosis to remote monitoring The most consequential translation has been in early detection of neurological disorders. A multi-center study from theGait and Brain Health Consortium(2025,The Lancet Digital Health) used a smartphone’s built-in accelerometer and gyroscope to collect 7-day continuous gait data from 8,200 older adults. A convolutional autoencoder trained on stride-to-stride variability identified a unique “fragile gait signature” characterized by increased phase-dependent entropy, which preceded clinical diagnosis of mild cognitive impairment by an average of 18 months (hazard ratio 2.4, 95% CI 1.8–3.1). This suggests that gait is not merely a motor output but a sensitive window into central nervous system integrity. Concurrently, in orthopedics,Tanaka et al.(2024,Clinical Biomechanics) demonstrated that IMU-derived foot strike angle and vertical loading rate, measured during a 10-meter walk in a clinic corridor, could predict prosthetic socket revision within 12 months for transtibial amputees with 87% accuracy—surpassing traditional radiographic assessments.

6. Remaining challenges: Heterogeneity, explainability, and workflow integration Despite these advances, three critical barriers persist. First,data heterogeneity: models trained on one population (e.g., healthy young adults) degrade significantly when applied to older patients with comorbidities, due to differences in cadence, body mass index, and footwear. Transfer learning and domain adaptation techniques are emerging, but they require large annotated datasets across diverse demographics—an expensive and ethically sensitive endeavor. Second,explainability: deep learning models often act as black boxes. Clinicians require not just a prediction (e.g., “fall risk: 75%”) but a mechanistic explanation (e.g., “increased double support time and reduced ankle push-off power”). While SHAP values and attention maps provide partial insights, they do not align with biomechanical terminology. A promising direction is the use ofconcept bottleneck models, where the network is forced to predict interpretable intermediate variables (e.g., joint angles, ground reaction forces) before making a final clinical decision. Third,workflow integration: most gait analysis software remains research-grade, lacking interoperability with electronic health records (EHRs) and reimbursement codes. A 2025 survey of 250 physical therapists found that only 18% used quantitative gait analysis in routine practice, citing time burden and lack of training as primary barriers.

7. Future outlook: Toward continuous, passive gait monitoring The next five years will likely see the convergence of three technologies: (i) ultra-low-power edge AI chips that process IMU data on-device, eliminating the need for cloud transmission and preserving battery life for weeks; (ii) smart textiles with embedded stretchable sensors that measure not only kinematics but also plantar pressure and shear forces, without the discomfort of rigid IMUs; and (iii) federated learning frameworks that train models across multiple hospitals without sharing raw patient data, addressing privacy concerns. We anticipate that gait will become a routine “vital sign” in telemedicine, analogous to heart rate or blood pressure—collected passively during daily activities and analyzed via a standardizedGait Health Indexthat combines speed, symmetry, variability, and energy expenditure. However, this future requires a radical rethinking of regulatory frameworks: how do we validate an AI model that is continuously updated with real-world data? Adaptive clinical trials and continuous performance monitoring will be essential.

Conclusion Gait analysis is undergoing a renaissance, propelled by advances in sensing hardware, deep learning, and biomechanical simulation. The integration of markerless vision, IMU-EMG fusion, and digital twin technologies has transformed gait from a coarse clinical observation into a precise, predictive, and personalized tool. Yet, the ultimate success will be measured not by algorithmic accuracy but by its ability to change clinical decisions and improve patient outcomes. As we move toward continuous, home-based monitoring, gait analysis holds the promise of detecting disease earlier, titrating treatments more precisely, and restoring mobility with greater confidence.

References