Embedded Neural Networks for Continuous Patient Posture Classification

Abstract

Current hospital protocols dictate patients be turned at least every three hours in the effort of preventing pressure ulcers. To reduce the workload of nurses, Momo Medical has created an embedded sensing device to track the patient's posture and notify nurses when it is time to turn them. The challenge presented and the focus of this thesis is classifying the posture of the patient based on the sensor data sampled, specifically, utilizing neural networks on an embedded platform.
Furthermore an optimization inspired by recurrent neural networks and ensemble neural networks is proposed, implemented, and compared against vanilla neural networks and pruned variants.