One Hip Wonder: 1D-CNNs Reduce Sensor Requirements for Everyday Gait Analysis

Abstract

Wearable inertial measurement units (IMU) enable large-scale multicenter studies of everyday gait analysis in patients with rare neurodegenerative diseases such as cerebellar ataxia. To date, the quantity of sensors used in such studies has involved a trade-off between data quality and clinical feasibility. Here, we apply machine learning techniques to potentially reduce the number of sensors required for real-life gait analysis from three sensors to a single sensor on the hip. We trained 1D-CNNs on constrained walking data from individuals with cerebellar ataxia and healthy controls to generate synthetic foot data and predict gait features from a single sensor and tested them in free walking conditions, including the everyday life of unseen subjects. We compare 14 stride-based gait features (e.g. stride length) with three sensors (two on the feet and one on the hip) with our approach estimating the same features based on raw IMU-data from a single sensor placed on the hip. Leveraging layer-wise relevance propagation (LRP) and transfer learning, we determine driving elements of the input signals to predict individuals’ gait features. Our approach achieved a relative error ($< 5\%$) similar to the state of the art three-sensor approach. Thus, machine learning-assisted one-sensor systems can reduce the complexity and cost of gait analysis in upcoming clinical studies while maintaining clinical meaningful effect sizes.