Ecological Validation of Machine Learning Models for Spatiotemporal Gait Analysis in Free-Living Environments Using Instrumented Insoles

Continuous gait monitoring may aid the early diagnosis of neurological and musculoskeletal conditions and help validate the effectiveness of new treatments. However, consumer-grade activity trackers can only capture summary gait metrics, whereas most of the research-grade devices capable of estimating fine-grained gait parameters are too cumbersome for extended-time use in real-life environments. Instrumented footwear may offer a promising alternative tool owing to their good accuracy and relatively small form factor, but their ability to detect stride-by-stride spatial gait parameters in free-living conditions has not been well explored to date. This work describes machine learning (ML) inference models for an insole system capable of accurately estimating stride time (ST), length (SL), and velocity (SV) in real-life environments. Functional validity was assessed through unstructured tests including straight-line walking, curve walking, and turns. Ecological validity was examined in free-living conditions. The ML models demonstrated better accuracy than conventional data processing methods (mean absolute errors in unstructured conditions were 3.55% for SL and 3.59% for SV). Real-life gait parameters estimated with the ML models showed stronger associations with a standardized walking test compared with the same parameters obtained with conventional methods. These results indicate proof-of-concept feasibility of using instrumented insoles and ML inference models for free-living spatiotemporal gait analysis.