Paper Title
Evaluation of Pre-Impact Fall Prediction Model using Machine Learning-Based Classifiers

Abstract
In US, over 70% of elders visit emergency room caused by fall-related injury. To reduce the chance of falls, pre-impact fall prediction model is developed. It detects falls before body hits on ground which will reduce long lie by sending alarm immediately or prevent or reduce injuries by integrating with on demand fall protective system. Previous studies used machine learning algorithm for pre-impact fall prediction model to overcome the limitation of threshold-based methods. However, the machine learning based fall prediction was classified as binary classes which do not consider various types of falls and activity daily livings (ADLs). Additionally, most of studies used window-based data for classification model instead of time-series based datasets. Therefore, in this study, the machine learning based fall prediction model is developed to predict multi-class fall prediction model which can classify 10 different types falls along with 14 types of ADLs and evaluated the performance of 4 different machine learning classifiers along with window based versus time-series based models. Our results showed that the long short-term memory showed highest average classification accuracy of 99.946 ± 0.072 % with the number of false negative and positive misclassification of 15, respectively. Therefore,the developed pre-impact fall prediction model can send the injury details of patients for better emergency treatment or allow the fall protective or exoskeleton systems to perform more precise feedback mechanisms which may reduce the incidence of falls and severity of physical injuries. Keywords - Pre-impact fall prediction, Inertial measurement unit, Lead time, Machine learning