Predicting ground contact events for a continuum of gait types: An application of targeted machine learning using principal component analysis

- A novel method to detect foot strike and toe-off has been developed for both walking and running.
- The method is shown to have a 95% prediction interval of ±20 ms from gold standard events.
- This method can potentially detect events using any measurement system that provides joint angles.

An ongoing challenge in the application of gait analysis to clinical settings is the standardized detection of temporal events, with unobtrusive and cost-effective equipment, for a wide range of gait types. The purpose of the current study was to investigate a targeted machine learning approach for the prediction of timing for foot strike (or initial contact) and toe-off, using only kinematics for walking, forefoot running, and heel-toe running. Data were categorized by gait type and split into a training set (∼30%) and a validation set (∼70%). A principal component analysis was performed, and separate linear models were trained and validated for foot strike and toe-off, using ground reaction force data as a gold-standard for event timing. Results indicate the model predicted both foot strike and toe-off timing to within 20 ms of the gold-standard for more than 95% of cases in walking and running gaits. The machine learning approach continues to provide robust timing predictions for clinical use, and may offer a flexible methodology to handle new events and gait types.