A fully embedded AI system for the detection of soft falls using accelerometer data in real time

Fall detection has been widely studied given the incidence of falls in the elderly population, often resulting in physical injuries that require prolonged medical treatment. This paper proposes a detection system to detect falls, and more importantly "soft falls". Soft falls do not necessarily produce physical injuries, but may be the consequence of a serious health issue such as a heart attack. Soft falls are more difficult to detect as they can be mistaken for daily activities. The detection system implements an artificial neuronal network model that reaches an AUC score of 98%. The model is reduced in size using a TinyML framework to be deployed in real time on a microcontroller, making it compatible with wearable systems that do not invade or restrict mobility and activities of users.