An Analysis of Segmentation Approaches and Window Sizes in Wearable-based Critical Fall Detection Systems with Machine Learning Models

Falls are the major risks and threats among elderly population. Various studies have developed automatic critical fall detection systems for emergency alarms and medical services. Window sizes and segmentation approaches would affect the system performance in terms of power consumption, computational speed and reliability that are essential considerations in the real-world implementation. Intuitively, employing shorter window sizes and simpler segmentation approaches leads to a faster detection speed and better power efficiency. However, few works explore the effects of windowing methods on detection performance, especially for machine learning models. This paper investigates the impact of segmentation approaches and window sizes in wearable-based critical fall detection systems with machine learning models. Two generally used segmentation approaches, including sliding windows and impact-defined windows, are explored with a range of window sizes from 0.5 s to 5.0 s and four machine learning models: Support Vector Machine (SVM), k-Nearest Neighbor (k-NN), Classification and Regression Tree (CART) and Naive Bayes (NB). The results show that sliding windows are more sensitive to the changes of window sizes compared to impact-defined windows. However, the differences between the best performance of both segmentation approaches are within 1% (99.66% in sliding windows and 99.78% in impact-defined windows). For systems with SVM and k-NN models, using 0.5 s window sizes in sliding windows and combinations of 1.5 s backward and forward windows in impact-defined windows are sufficient to achieve at least 94% accuracy and 98% accuracy.