AI in Sleep Medicine

The standard sleep study is polysomnography (PSG), an overnight recording of brain waves (EEG), eye movements, chin muscle activity, heart rhythm, breathing airflow, chest and belly effort, and blood oxygen. Technologists divide the night into 30-second epochs and label each one as wake, N1, N2, N3 or REM, following the American Academy of Sleep Medicine scoring rules. They also mark apneas (breathing nearly stops for at least 10 seconds), hypopneas (partial reductions linked to oxygen drops or arousals), arousals and leg movements. The apnea-hypopnea index (AHI) is the number of apneas and hypopneas per hour of sleep. In adults, 5 to 15 is usually called mild, 15 to 30 moderate, and 30 or more severe. Scoring a full night by hand is slow, and human scorers do not agree perfectly. Agreement is lowest for stage N1. AI autoscoring models, often deep neural networks trained on thousands of scored studies, can stage sleep and detect events in minutes. Several commercial products have FDA clearance as aids whose output a qualified person reviews. Home sleep apnea tests use fewer sensors, and some devices use signals such as peripheral arterial tone from the finger, with algorithms estimating sleep time and respiratory events. Consumer watches from Samsung and Apple have received FDA authorization for sleep apnea notification features based on signals such as motion or blood oxygen patterns. They are screening prompts, not diagnoses. The main misconception is that a wearable sleep score equals a sleep study. Wearables usually estimate stages from movement and heart rate rather than EEG, often mistake quiet wakefulness for sleep, and are less reliable in people with insomnia, neurological conditions or irregular heart rhythms.