Can machine learning techniques reduce the number of inertial sensors in real life gait analysis?
BACKGROUND AND AIM: The optimal number of inertial sensors for real-life gait analysis is a trade-off between data quality and patient convenience and feasibility. One-sensor systems have proven to deliver reliable information for average values of gait speed or step length. However, for the ataxic-sensitive measures of spatio-temporal gait variability, these systems reported less reliability and less sensitivity compared to 3 sensor systems including two sensors at the feet. Here, we investigate the potential of machine learning techniques to predict gait features based on 1 sensor only, which could increase the clinical feasibility of instrumented gait analysis in real-life recordings of cerebellar ataxic patients. METHODS: We recorded gait data from 44 healthy controls and 55 cerebellar patients at baseline, 1-year and 2-years follow-up assessments by 3 Opal APDM inertial sensors. These data successful identified longitudinal changes in gait variability measures for cerebellar patients (e.g. stride length variability, effect size: 0.53) Utilising 1D convolutional neural networks (1D-CNN) we predicted 14 gait parameters from stride based triaxial IMU data in two conditions with different input dimensions: using raw data from the pelvis sensor only (1S) in comparison to the complete set of all three sensors (3S). Thus, in the supervised training phase of both conditions, we used stride based gait features previously determined by the 3 sensors algorithm from APDM as ground truth. Aim in both approaches is to individualize the learned mappings for a new unseen patient based on a small amount of recorded gait samples with 3 sensors in the lab and to use transfer learning for the characterisation of real-life data. RESULTS: First results deliver a low (