Phase 1: Fall mechanics study In the lab setting, hundreds of controlled falls on different weighted mannequins were performed from sitting, standing and higher altitude positions. Using FallCapture to record accelerometer data on Apple Watch, the fall types were marked and thousands of data points were obtained regarding the different phases of a fall event. Following an extensive analysis, it was observed that lower falls from a sitting position produced different signatures compared to falls from a standing position. From this, the smart fall detection algorithm was born.
Phase 2: Proof of concept lab study. Low vs. high fall sensitivity and accuracy determination. From the original data, prototype, app-based versions of FallCall Detect were built. All parts of the fall detection algorithm were set and controlled fall testing was conducted. After hundreds more falls from the sitting and standing positions were conducted, it was observed that over 80% of falls could be detected with average high vs. low accuracy determination ranging from 60% + to 80%+. From additional experiments, our team defined algorithm adjustments that were necessary to minimize false activations while still being able to detect and differentiate falls.
Phase 3. Final validation studies. With optimization studies completed, the FallCall Detect team spent months validating its fall detection algorithm. Thousands of fall events were recorded using mannequin and human subjects wearing Apple Watches on each wrist. Various hand/arm locations were tested from the sitting and standing positions. Fall detection was able to capture up to 90%+ of higher impact standing falls and up to 80% of lower impact sitting falls with up to 70% differentiation accuracy depending on fall mechanics. Additionally, human subjects wore Apple Watches daily to determine false activation rates based on normal activity and exercise activities. Using daily usage data combined with lab data, adjustable sensitivities were determined to maximize fall detection sensitivity and fall differentiation accuracy while minimizing false activations.