TRUSTWORTHY SLEEP STAGING FROM EEG: DEEP ENSEMBLES, MCDROPOUT, AND PREDICTIVE CALIBRATION
Reliable sleep stage classification from EEG signals is critical for the development of clinical decision support systems. However, many deep learning models lack mechanisms for estimating predictive uncertainty, which is important for trust and interpretability. This work explores the use of Monte Carlo Dropout and Deep Ensembles to estimate uncertainty in automatic sleep staging. We apply these methods to convolutional and recurrent neural architectures trained on a publicly available EEG dataset. Evaluation includes calibration analysis and visualization of predictive entropy across sleep stages. LSTM-based models demonstrated more consistent trustworthiness and calibration across runs, while EEGNet remained appealing for lightweight deployment.The results highlight the value of incorporating uncertainty estimates to improve model transparency and support safe use in medical contexts.