Validation of a Deep Learning-Based Markerless Video System for Gait Speed Assessment During the 10-Meter Walk Test in Outdoor Clinical Settings

Abstract

Background/Objectives: Markerless video-based gait assessment has emerged as a promising alternative to laboratory motion capture systems because it enables objective, low-cost, and accessible mobility assessment. However, evidence supporting the validity of deep learning (DL)-based markerless video systems for gait speed assessment during the 10-meter walk test (10-MWT) in real-world clinical environments remains limited. This study evaluated the concurrent validity of a DL-based markerless video system for estimating gait speed during the 10-MWT under outdoor clinical conditions. Methods: Thirty-two healthy participants from three age groups completed the 10-MWT under comfortable, slow, and fast walking conditions. Gait speed was simultaneously measured using the proposed DL-based markerless video system with the YOLOv11x model and Tracker video analysis software (version 6.3.2). Gait speed was calculated across six consecutive 1 m walking segments. Concurrent validity was evaluated using Pearson’s correlation coefficients and Bland-Altman analysis. Results: Pearson’s correlation coefficients ranged from 0.92 to 0.98 across all six walking segments and walking-speed conditions, indicating very high concurrent validity. Bland-Altman analysis demonstrated good agreement between the proposed system and the reference method, with mean differences close to zero and most observations falling within the 95% limits of agreement. The proposed system maintained high agreement across comfortable, slow, and fast walking conditions while enabling segment-specific gait speed analysis throughout the 10-MWT. Conclusions: The proposed DL-based markerless video system demonstrated high agreement with Tracker for gait speed assessment during the 10-MWT under outdoor non-laboratory conditions, supporting its potential as an accessible approach for gait speed assessment.