Robust Deep Learning Model for Accurate Fall Detection using Smartphone Sensor Data

Abstract

Falls and associated medical conditions are a serious concern in healthcare, especially among elderly people. The ratio of older people is growing every year, making it even more crucial to have effective fall prevention and detection systems in place. The main objective of this research is to provide a comprehensive solution from dataset creation to application of effective modeling techniques in the context of Fall Detection systems (FDS). A new dataset is presented based on scripted Activities of Daily Life (ADL) performed by elderly people. This representative data combined with young volunteer’s simulated falls data is the key differentiation to offer a relevant basis for prediction. The proposed work intends to develop a binary classification framework that can analyze the data and correctly categorize falls and no falls by distinguishing falls from complex fall like activities of daily life (ADL). Recurrent Neural Networks (RNNs) having the ability to handle sequential data and capture temporal dependencies are used in this research for robust fall detection based on smartphone accelerometer data. The accuracy of FDS is most of the time assessed by the performance metrics only, without performing organized testing on varied set of data. Adequate strategies along with thorough testing across several datasets is imperative for a reliable FDS. To add to the sophistication of this research the employed fall detection techniques are assessed using a variety of public datasets. Proposed models are evaluated for the reference new dataset and two other publicly available datasets. With all these considerations the proposed approach and model has shown better performance as compared to previously adopted models. Keywords: Accelerometer, fall detection, Deep Learning, smart phone data, Random Forest, BiLSTM, feature selection, Machine Learning