Fall risk assessment through a synergistic multi-source DNN learning model

Abstract

Falls are a complex problem and play a leading role in the development of disabilities in the older population. While fall detection systems are important, it is also essential to work on fall preventive strategies, which will have the most significant impact in reducing disability in the elderly. In this work, we explore a prospective cohort study, specifically designed for examining novel risk factors for falls in community-living older adults. Various types of data were acquired that are common for real-world applications. Learning from multiple data sources often leads to more valuable findings than any of the data sources can provide alone. However, simply merging features from disparate datasets usually will not produce a synergy effect. Hence, it becomes crucial to properly manage the synergy, complementarity, and conflicts that arise in multi-source learning. In this work, we propose a multi-source learning approach called the Synergy LSTM model, which exploits complementarity among textual fall descriptions together with people's physical characteristics. We further use the learned complementarities to evaluate fall risk factors present in the data. Experiment results show that our Synergy LSTM model can significantly improve classification performance and capture meaningful relations between data from multiple sources.