Accurate sleep monitoring remains a challenge for consumer wearables, and conditions like Obstructive Sleep Apnea are widely underdiagnosed due to the lack of accessible, reliable automated tools. Using the DREAMT dataset — combining overnight PSG signals, Empatica E4 smartwatch data, and subject metadata from 100 participants — we built two models: one to classify sleep stages in real time, and one to predict apnea events 10 seconds in advance.
For sleep stage classification, we benchmarked Logistic Regression, XGBoost, LightGBM, and LSTM; gradient boosting methods performed best, with an XGBoost and LightGBM ensemble further improved by majority-vote smoothing. For apnea prediction, a longer lag window of LightGBM features yielded the strongest results, highlighting the importance of temporal context.
These models enabled real-time stage tracking and proactive apnea alerts with potential for earlier clinical interventions.