A machine learning model to detect falls mimicking cardiac arrest-related collapse based on wrist-derived...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning in Healthcare

Background:

• Photoplethysmography (PPG) is common in automated cardiac arrest detection but prone to false positives.
• Accelerometry signals are used to detect movement and reduce false alarms in wearable devices.

Purpose of the Study:

• To develop and evaluate an accelerometer-based machine learning model for detecting cardiac arrest-related collapse.
• To assess the model's sensitivity and false-positive rate in identifying simulated falls.

Main Methods:

• The DETECT-2 study used accelerometer data from the CardioWatch wristband.
• A gradient boosting model (GBM) was trained and tested on simulated cardiac arrest-related collapses (sudden and soft falls).
• Video recordings served as a reference for fall detection.

Main Results:

• The GBM achieved high sensitivity in detecting falls: 99.2% in the training set and 96.1% in the test set.
• Sudden falls were detected with 100% sensitivity, while soft falls had 88.1% sensitivity in the test set.
• The model demonstrated a low false-positive rate, with only four in the training set and two in the test set.

Conclusions:

• Accelerometer data can accurately detect falls mimicking cardiac arrest-related collapse.
• Integrating accelerometer signals into PPG-based models can reduce false positives and enhance sensitivity for automated cardiac arrest detection.