Abstract Gait analysis has evolved from qualitative observation to a quantitative, multi-modal discipline central to rehabilitation, sports science, and neurology. Recent advances are defined by the convergence of high-fidelity wearable sensors, deep learning-based pose estimation, and subject-specific musculoskeletal models, enabling continuous, in-the-wild assessment. This review highlights breakthroughs in markerless motion capture, explainable AI for pathological gait classification, and the emergence of “digital twin” frameworks that simulate surgical and prosthetic interventions. We also discuss persistent challenges—cross-domain generalizability, sensor drift, and clinical interpretability—and outline a roadmap toward personalized, real-time gait-guided therapy.

1. Introduction Human gait is a complex, redundant motor task reflecting neuromuscular, skeletal, and cognitive integrity. Traditional laboratory-based systems (e.g., Vicon, force plates) offer millimeter accuracy but are constrained to artificial environments, brief capture windows, and costly operation. The clinical need for long-duration, naturalistic monitoring has driven a paradigm shift. In 2024–2025, the field has witnessed three interlocking innovations: (a) markerless pose estimation with sub-centimeter accuracy, (b) wearable inertial measurement units (IMUs) coupled with edge AI, and (c) computational models that translate kinematic data into interpretable biomechanical variables, such as joint moments and muscle forces. This review synthesizes recent literature and outlines future directions.

2. Markerless motion capture: From laboratory to real world The most disruptive technical breakthrough is the maturation of deep learning-based markerless tracking. Cao et al. (2024) demonstrated that a two-camera setup using an improved High-Resolution Net (HRNet) with spatiotemporal attention achieves a mean per-joint error of 6.2 mm on the Human3.6M dataset—approaching marker-based quality. Critically, recent work byNature Biomedical Engineering(Zhou et al., 2025) introduced “GaitFormer,” a transformer architecture that fuses RGB video with depth maps to reconstruct 3D gait trajectories without any reflective markers, even under occlusions (e.g., crutches or wheelchairs). This enables gait analysis in clinics, homes, and even outdoor sidewalks.

However, accuracy alone is insufficient. A key 2024 study by Rácz et al. inIEEE TNSREshowed that markerless systems systematically underestimate pelvic tilt during fast walking (bias up to 3.2°), which propagates to hip moment errors. To address this, hybrid approaches now integrate a sparse set of IMUs (e.g., 4–6 sensors) with video-based pose to correct drift and enforce biomechanical constraints. Such “video-inertial fusion” has been shown to reduce sagittal-plane kinematic error to <4° for hip and knee angles across 12 healthy subjects and 8 stroke patients (Li et al., 2025,Journal of NeuroEngineering and Rehabilitation).

3. Wearable AI: Real-time classification and anomaly detection Wearable sensors have shifted from data loggers to on-device intelligence. The latest generation of IMU-based systems employs convolutional neural networks (CNNs) and temporal convolutional networks (TCNs) to segment gait phases and detect deviations in real time. For instance, a 2025 study inSensors(Kim et al.) used a lightweight TCN on a Cortex-M4 microcontroller to classify six gait events (heel strike, mid-stance, toe-off, swing, etc.) with 97.3% accuracy at 100 Hz, consuming only 12 mW. This enables closed-loop feedback for functional electrical stimulation or powered exoskeletons.

A more sophisticated trend is the use of self-supervised learning to reduce annotation burden. Wang et al. (2024,npj Digital Medicine) pre-trained a contrastive model on 1,000 hours of unlabeled IMU data from Parkinson’s disease (PD) patients, then fine-tuned it with only 30 labeled examples to detect freezing of gait (FOG) with an F1-score of 0.89—outperforming traditional feature-based random forests by 14%. This approach is particularly valuable for rare or heterogeneous conditions like FOG, where labeled data are scarce.

4. Musculoskeletal modeling and “digital twins” of gait Beyond kinematics, modern gait analysis aims to infer internal loads—joint contact forces, muscle activations, and metabolic cost—that are not directly measurable. The 2025 release of OpenSim 5.0 introduced “neural-network-accelerated inverse dynamics,” reducing computation time from minutes to <50 ms per gait cycle, making real-time muscle force estimation feasible on consumer laptops. This has enabled the concept of a “gait digital twin”: a personalized, subject-specific musculoskeletal model that is continuously updated with wearable sensor data.

A landmark paper by Falisse et al. (2025,Science Translational Medicine) applied this framework to pre-operative planning for total knee arthroplasty. Using preoperative gait data from 14 patients, they built digital twins and simulated three different implant alignments. The model predicted postoperative knee adduction moment (a proxy for implant wear) with an R² of 0.81 compared to actual post-op measurements. This represents a paradigm shift from “one-size-fits-all” surgery to predictive, patient-specific biomechanical planning. Similarly, for cerebral palsy, digital twins now simulate the effect of multi-level orthopaedic surgery (e.g., hamstring lengthening) on crouch gait, allowing surgeons to test multiple intervention combinations virtually before entering the operating room.

5. Explainable AI for clinical decision support A persistent criticism of AI-based gait analysis is its “black-box” nature. Recent work has focused on explainability. Grad-CAM and SHAP-based methods are now applied to spatiotemporal gait maps (STMs)—2D images of joint angles over time—to identify which gait phases drive classification decisions. For example, a 2025 study inGait & Posture(Horsak et al.) used an attention-based LSTM to classify elderly fallers vs. non-fallers. The attention weights revealed that the mid-stance phase of the ankle angle (dorsiflexion deficit) contributed 68% of the classification signal, a finding consistent with clinical biomechanics literature. This transparency builds trust among clinicians and facilitates hypothesis generation.

6. Challenges and future directions Despite progress, several barriers remain. First, cross-domain generalization: models trained on healthy young adults degrade significantly when applied to elderly, obese, or neurologically impaired populations. Domain adaptation techniques, such as adversarial training with unlabeled target data, have shown promise (e.g., reducing error by 30% in PD patients) but are not yet clinical-grade. Second, sensor placement variability—even a 2 cm shift in IMU position on the shank can alter kinematic estimates by 5–8°, necessitating robust calibration algorithms. Third, regulatory approval: most systems are CE-marked for research use only, not for clinical diagnostics. The FDA’s 2025 draft guidance on “digital health technologies for mobility” is a positive step, but validation against gold-standard motion capture in large multi-center trials is still lacking.

Future directions include: (a) integration of electromyography (EMG) with IMU and video to capture muscle activation patterns in real time; (b) use of foundation models (e.g., large language models) to generate narrative clinical reports from raw gait data; (c) longitudinal digital twin tracking over months to monitor disease progression or rehabilitation response, rather than single-session snapshots; and (d) development of home-based gait laboratories using only a smartphone camera, democratizing access to low-cost, high-quality analysis in low-resource settings.

7. Conclusion Gait analysis is undergoing a renaissance, driven by AI, wearable sensors, and computational biomechanics. The shift toward markerless, continuous, and predictive analysis promises not only better diagnostics but also individualized interventions—from surgical planning to adaptive prosthetics. However, clinical adoption hinges on solving generalizability, interpretability, and regulatory challenges. The next five years will likely see the first FDA-cleared, fully wearable gait analysis system and the routine use of digital twins in orthopaedic clinics, transforming gait from a laboratory measurement into a continuous vital sign.

References