Validation of AI-based markerless gait event detection during perturbed walking using smartphone videos from two...

Perturbation-based balance training (PBT) requires objective quantification of reactive stepping responses, yet gait event detection during perturbed walking remains methodologically challenging. Markerless smartphone-based gait analysis offers a scalable alternative to laboratory motion capture systems, but its robustness under perturbation-induced gait irregularities is insufficiently validated. This pilot study including five participants explored the validity of the SMARTGAIT markerless motion analysis system during normal and perturbed treadmill walking in older adults and investigated whether perturbation-specific model retraining could improve gait event detection.

Five geriatric participants (mean age 83 ± 4.47 years; 40% female) walked on a perturbation treadmill while eight unexpected perturbations in both mediolateral and anteroposterior displacement were administered. Smartphone videos from frontal and diagonal perspectives were recorded and gait events (initial contact, final contact) were manually annotated as ground truth. Model performance was evaluated using F1 score, precision, and recall with a ± 33 ms tolerance window. Retraining was conducted using five-fold cross-validation on perturbed segments.

Across 2,477 annotated gait events (2,106 normal; 371 perturbed), initial detection performance was lower during perturbed walking (median F1 score: frontal 0.49; diagonal 0.88) compared to normal walking (median F1 score: frontal 0.68; diagonal 0.96), with large effect sizes despite non-significant p-values (frontal: p = 0.10; diagonal: p = 0.06). This decrease was mainly recall-driven, whereas precision remained high. No statistically significant differences were found between diagonal and frontal camera perspectives; however, large effect sizes (normal walking: r = 0.87; perturbed walking: r = 1.00) suggest that camera perspective may influence detection accuracy. Retraining was associated with higher perturbed gait event detection in the frontal perspective (F1 score: 0.49 → 0.76, p = 0.06), mainly accompanied by higher recall (0.34 → 0.64, p = 0.06), but showed no change in the diagonal perspective (F1 score: 0.88 → 0.88; p = 0.18).

This pilot study provides proof of concept for the application of smartphone-based markerless gait event detection during perturbation-induced walking. The findings further support perturbation-specific model retraining as a feasible strategy for adapting AI-based gait analysis to challenging movement conditions. Larger validation studies are required to confirm these preliminary findings before clinical implementation.