Abstract. Wearable inertial measurement units (IMU) enable largescale 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 (