Supervised Machine Learning-based Fall Detection

Falls are a serious public health problem, and in the recent years, the number of proposed fall detection systems that has been developed increased dramatically. In this study, a supervised machine learning-based fall detection was developed using a wearable accelerometer device fitted to the subject's waist. Fifteen volunteers performed a standardized set of movements which were developed by mimicking the scenarios in the FARSEEING Fall Repository, including 10 falling activities and 10 non-falling activities, resulting in a dataset with 900 samples. Supervised Learning is implemented by using two training sets: (1) 80% of the total simulated fall and non-fall activities and (2) 10 real-world fall signals from the FARSEEING Fall Repository added to the 80% of the total simulated fall and non-fall activities. Testing is done by using the remaining 20% of the total simulated fall and non-fall activities, remaining five real-world fall signals from the FARSEEING Fall Repository, and actual testing in a simulated environment. Results show that the model with the training data using simulated falls only (Model 1) detected a 10% TPR while the model with real-world fall added (Model 2) detected a 100% TPR when tested using unseen data. Validating using the actual simulated fall and non-falls, Model l resulted in a 93.33% TPR and Model 2 at 96.67%. Hence this shows that the second training set performed better in classifying simulated fall, non-fall, and real-world fall signals as a fall or non-fall.