Leveraging ML on Biometric Sensor Data to Detect and Predict Sleep Activity

Title: Leveraging ML on Biometric Sensor Data to Detect and Predict Sleep Activity

Author: Bronson Tharpe

Faculty Sponsor: Dr. Anu Bourgeois, Associate Professor, Department of Computer Science

Motivation: Wearable devices such as fitness trackers and smart watches hold tremendous medical importance in the modern world. Previous research has demonstrated that machine learning has the potential to save lives in concordance with such medical technology. We set out to harness predictive measurements from Fitbit Charge 4 and Samsung Galaxy Watch Active devices and utilize them for sleep prediction.

Methods:. Using a crowd-sourced Zenodo database, we extract novel features from simple biometrics such as heart-rate and step count and use these data to train a custom Gradient Boosting Classifier model. This model demonstrates greater than 95% accuracy for sleep detection and over 92% accuracy at predicting whether a user will be asleep within the next 30 minutes. Although our model was finely tuned for sleep prediction, we offer a robust framework for anticipating a range of other biological events. Thus, our solution provides the scaffolding for future predictive technologies that can result in significant human benefit.

Discussion of Importance: We hope for future research to focus on the prediction and prevention of harmful events, such as seizures and heart attacks. Our final model provides a basis for generalizing to other biological events.