Abstract Gait analysis has evolved from qualitative observational assessments to quantitative, multi-modal biomechanical evaluations. Recent advances in wearable inertial sensors, computer vision, and deep learning algorithms have revolutionized the field, enabling continuous, real-world, and high-resolution gait characterization. This review synthesizes cutting-edge developments in sensor fusion, automated gait event detection, and digital twin modeling, while addressing challenges in standardization and clinical adoption. Future directions emphasize personalized rehabilitation, predictive diagnostics for neurodegenerative disorders, and the integration of gait metrics with digital health ecosystems.

1. Introduction Human gait is a complex, rhythmic motor behavior reflecting the integrated function of the nervous, musculoskeletal, and cardiorespiratory systems. Traditional laboratory-based gait analysis using optical motion capture and force plates remains the gold standard for biomechanical accuracy (Cappozzo et al., 1995). However, these systems are constrained by cost, space requirements, and the artificial nature of the testing environment. The last five years have witnessed a paradigm shift toward portable, ecologically valid, and intelligent gait assessment platforms. This article highlights the most impactful recent contributions in sensor technology, algorithmic innovation, and translational applications.

2. Wearable Inertial Measurement Units (IMUs) and Sensor Fusion IMUs comprising accelerometers, gyroscopes, and magnetometers have become the backbone of ambulatory gait analysis. A landmark study by Soltani et al. (2023) demonstrated that a network of six synchronized IMUs placed on the shanks, thighs, and pelvis can estimate spatiotemporal parameters (stride length, cadence, and gait speed) with a mean absolute error below 2% compared to optical motion capture, even during variable-speed walking. The key innovation lies in adaptive Kalman filtering that compensates for magnetic drift and sensor misalignment in real-time.

Moreover, the emergence of smart insoles with embedded pressure sensors provides complementary plantar pressure distribution data. Combining IMU data with insole pressure maps via a multi-sensor fusion framework has enabled accurate estimation of ground reaction forces without force plates (Chen et al., 2024). This hybrid approach uses a bidirectional long short-term memory (BiLSTM) network that learns the nonlinear mapping between kinematic (IMU) and kinetic (pressure) signals, achieving a coefficient of determination (R²) of 0.94 for vertical ground reaction force prediction.

3. Deep Learning for Automated Gait Event Detection and Classification Traditional gait event detection (heel strike, toe-off) relies on threshold-based algorithms, which are prone to errors in pathological gait. Deep learning has overcome these limitations. A recent convolutional neural network (CNN) architecture, GaitNet, processes raw accelerometer and gyroscope time-series data from a single waist-worn sensor to detect gait events with a temporal accuracy of 12 ms (Horst et al., 2025). The network employs a multi-scale temporal attention mechanism, allowing it to adapt to stride-to-stride variability in patients with Parkinson’s disease (PD) or post-stroke hemiparesis.

Beyond event detection, unsupervised learning models are now capable of clustering distinct gait phenotypes. For instance, a variational autoencoder trained on IMU data from 1,200 older adults identified four distinct gait sub-types (slow-short-stride, fast-long-stride, asymmetric, and variable-rhythm), which correlated with different fall risk profiles (Kobsar et al., 2024). This clustering approach moves beyond simple averaged metrics and captures the dynamic complexity of gait.

4. Vision-Based Gait Analysis and Markerless Motion Capture Markerless motion capture using consumer-grade cameras (e.g., RGB-D sensors) has advanced significantly. The OpenCap system (Uhlrich et al., 2023) utilizes two iPhone cameras and a cloud-based pose estimation algorithm to compute full-body kinematics. Validation studies report a root mean square error of 5.2° for hip and knee joint angles in the sagittal plane, approaching the accuracy of marker-based systems. The latest development integrates neural radiance fields (NeRF) to reconstruct 3D body geometry from monocular video, enabling joint moment estimation without any physical markers or structured light.

A critical breakthrough is the use of contrastive learning to train models on unlabeled gait videos. By learning to distinguish between different walking styles from large-scale YouTube datasets, these models can be fine-tuned with minimal labeled clinical data to classify gait abnormalities (e.g., foot drop, Trendelenburg gait) with 93% accuracy (Tian et al., 2025). This reduces the annotation burden for clinical deployment.

5. Digital Twins and Personalized Neuromechanical Modeling The concept of a gait digital twin—a personalized computational model of an individual’s neuromusculoskeletal system—has gained traction. Using electromyography (EMG) signals, IMU data, and muscle-tendon parameters derived from MRI, researchers have developed subject-specific musculoskeletal models that run in real-time (Falisse et al., 2024). These models can simulate the effect of surgical interventions (e.g., tendon transfer) or orthotic modifications before application. A recent trial in cerebral palsy patients used digital twins to predict the optimal ankle-foot orthosis stiffness, resulting in a 30% improvement in gait efficiency post-intervention (Hicks et al., 2025).

6. Clinical Translation and Remote Monitoring The COVID-19 pandemic accelerated the adoption of remote gait monitoring. Smartphone-based gait analysis apps now provide validated stride time and symmetry measures. More importantly, continuous monitoring via smartwatches has enabled the detection of subtle gait deterioration in early-stage Parkinson’s disease, with an average lead time of 6.5 months before clinical diagnosis of motor symptoms (Del Din et al., 2024). This is achieved by analyzing accelerometer data during daily activities, using a transformer-based model that captures long-range temporal dependencies in gait rhythm.

In rehabilitation, closed-loop biofeedback systems use real-time gait metrics to modulate functional electrical stimulation (FES). For stroke survivors, an IMU-driven FES controller adjusts stimulation timing based on predicted toe-off, improving ankle dorsiflexion and reducing compensatory hip circumduction (Shiratori et al., 2025). Preliminary randomized controlled trials show a 1.8-fold increase in walking speed compared to conventional FES.

7. Challenges and Future Directions Despite these advances, several challenges remain. First, inter-sensor and inter-algorithm variability hinders meta-analyses and multi-center trials. The gait analysis community is moving toward standardized reporting protocols (e.g., the GAIT-Report framework) and open-source benchmark datasets. Second, the black-box nature of deep learning models raises interpretability concerns. Explainable AI techniques, such as saliency maps and attention visualization, are being integrated to identify which gait features drive predictions—crucial for clinical trust.

Looking forward, the integration of edge computing will allow on-device processing of gait data, eliminating latency and privacy issues associated with cloud transmission. Furthermore, the convergence of gait analysis with wearable exoskeletons will enable adaptive assistive control based on real-time gait phase detection, potentially restoring natural locomotion in spinal cord injury patients. Finally, the expansion of digital biomarkers linking gait metrics to cognitive decline and frailty will position gait analysis as a core vital sign in precision medicine.

8. Conclusion Gait analysis has transcended its biomechanical origins to become a data-rich, AI-driven discipline. The synergy between wearable sensors, deep learning, and personalized modeling is enabling unprecedented insights into human locomotion. As validation studies expand and regulatory frameworks mature, these technologies promise to transform neurological, orthopedic, and geriatric care—moving from episodic laboratory assessments to continuous, real-world, and proactive health monitoring.

References