Wearable stress research
This is the research I did for my capstone in the BASS Emerging Scholars program, a summer research program run by BASS Medical Group (paper, poster, code). I trained machine learning models on wrist-sensor data from adults in alcohol use disorder treatment, to see if they could tell when someone was about to report feeling stressed. I tested every model on people it had never seen in training, so it couldn't just memorize individuals.
The best model, a random forest looking at 5-minute windows, got an average AUC of 0.77 across 11 participants and beat chance for every one of them. It's a small dataset, so it shows the signal is there, not that it's ready to use in clinics.
