Accuracy of a wavelet-based fall detection approach using an accelerometer and a barometric pressure sensor

Goal: Falls are a major source of morbidity in older adults, and 50% of older adults who fall cannot rise independently after falling. Wearable sensor-based fall detection devices may assist in preventing long lies after falls. The goal of this study was to determine the accuracy of a novel wavelet-based approach to automatically detect falls based on accelerometer and barometric pressure sensor data. Methods: Participants (n=15) mimicked a range of falls, near falls, and activities of daily living (ADLs) while wearing accelerometer and barometric pressure sensors on the lower back, chest, wrists and thighs. The wavelet transform using pattern adapted wavelets was applied to detect falls from the sensor data. Results: In total, 525 trials (194 falls, 105 near-falls and 226 ADLs) were included in our analysis. When we applied the wavelet-based method on only accelerometer data, classification accuracies ranged from 82% to 96%, with the chest sensor providing the highest accuracy. Accuracy improved by 3.4% on average (p=0.041; SD=3.0%) when we also included the barometric pressure sensor data. The highest classification accuracies (of 98%) were achieved when we combined wavelet-based features and traditional statistical features in a multiphase fall detection model using machine learning. Conclusion: We show that the wavelet-based approach accurately distinguishes falls from near-falls and ADLs, and that it can be applied on wearable sensor data generated from various body locations. Additionally, we show that the accuracy of a wavelet-based fall detection system can be further improved by combining accelerometer and barometric pressure sensor data, and by incorporating wavelet and statistical features in a machine learning classification algorithm.