Objective: Falling is an important health maintenance issue for the elderly and people with movement disorders, strokes and multiple sclerosis. With the development of light, low-cost wearable technology, inertia-based fall detection has gained much attention. However, some large movements, such as jumping and postural changes, are frequently confounded with falls. For example, commonly used fall detection methods based on acceleration amplitude produce a large number of false alerts unless they are combined with post-fall posture identification. In this paper, we propose two new inertial parameters to improve the selectivity of threshold-based fall detection methods, and evaluate strategies to distinguish falls from other activities of daily life (ADLs).
Approach: We define two new inertial parameters, acceleration cubic-product-root magnitude (ACM) and angular velocity cubic-product-root magnitude (AVCM). Along with acceleration magnitude (AM), we test threshold-based fall detection methods based on single parameters and combinations. We collected inertial data on four types of simulated falls and eight types of ADLs from a study with 15 participants wearing a chest-mounted sensor with accelerometer and gyroscope. Two public datasets, UMAFall and Cognent Labs, were also included to evaluate fall detection methods.
Main results: We chose the detection threshold with 99% sensitivity and the best possible specificity. The hybrid of AM, ACM and AVCM method had a lower rate of misclassification than single-parameter methods. Leave-one-out cross-validation shows that the hybrid fall detection method can achieve both high specificity and high sensitivity.
Significance: Using multiple inertial parameters improves the specificity of fall detection.