Automatic Sleep Stage Scoring from EEG

Sleep clinics score every 30-second epoch of an overnight recording as wake, N1, N2, N3 or REM by hand, which takes hours per patient. This project scores the single EEG channel Fpz-Cz of real overnight recordings (Sleep-EDF Expanded, PhysioNet; 10 healthy adults, 20 nights) automatically. It compares threshold rules, logistic regression, the Random Forest of the specification, gradient boosting and a Random Forest with neighbouring epochs - all on 12 spectral features - with a small 1-D convolutional network on the raw signal, each with and without hidden-Markov- model smoothing, and the pre-trained YASA stager as an external reference. Every result is measured on people the models never saw. The study also shows how an epoch-random split overstates agreement, why N1 is the hardest stage and how well clinical sleep metrics (total sleep time, efficiency, WASO, REM latency) are reproduced. A web application scores nights or uploaded EDF files, shows the hypnogram and the EEG of every epoch, lets a sleep technologist correct epochs and records the sign-off, on five databases. Decision support and education only - not a medical device.

Category: Biomedical Engineering

Biomedical Engineering Project BM03 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.