Accuracy Comparison of ML-Based Fall Detection Algorithms Using Two Different Acceleration Derived Feature Vectors

Falls among the elderly are an important health issue which can lead to serious consequences. Unintentional falls in the elderly are the most common cause of nonfatal injuries. Different algorithms have been proposed for fall detection using vision based, ambient or inertial sensors. In this experiment, we used data collected from 16 subjects performing 15 tasks: 12 activities of daily living and 3 simulated falls. Subjects were equipped with a waist-mounted IMU to collect tri-axial accelerometer data during the activities. We derived features from accelerometer data on time-windowed identified activities and divided them into 2 different feature vectors (FV) that are used for fall detection algorithm. The selection was made based on the computational complexity of the features. We used different machine learning (ML) techniques to distinguish fall events from non-fall events (i.e. activities of daily living). In this paper, we compared and analyzed the accuracy of fall detection for different ML-based algorithms using the two accelerations-derived feature vectors for potential real-time implementation. The accuracy of the selected models ranged from 80.1 to 90.0% for FV1 and from 87.2 to 98.6% for FV2. The support vector machines algorithm achieved the best performance with accuracy 98.6%, sensitivity 98.7% and specificity 98.6% for FV2.