SVM-based model for extracting heart rate signal segments from existing or realtime PPG (photoplethysmogram) measurements, which usually tend to be difficult to parse, noisy, and motion affected, particularly when originating from a sensor embedded in a wearable device. The model can be integrated into a mobile application to automatically detect heart rate patterns in live PPG stream (e.g. via BLE) and extract meaningful data for visualization.

The training set, located in dataPPG/, is comprised of anonymized PPG readings with corresponding timestamps, taken using multiple wearable devices from children at a Montessori preschool over the course of one week. Readings were obtained with an NJR reflective sensor and the popular TI link medical FE.

Input data should be raw PPG readings, such as this:

Long term PPG measurement

Positively classified segment, zoomed in;
97.5 beats per minute