Approaching the Human Ceiling in Sleep Staging: A Controlled Comparison of Feature-Engineered and Deep Architectures

Abstract

Polysomnography (PSG) is the clinical gold standard for sleep assessment, yet manual epoch-by-epoch scoring is labor-intensive and subject to substantial inter-rater variability, which limits its use in large-scale, home-based, and routine clinical practice and makes reliable automated scoring an urgent need. This study suggests and compares an automated sleep-staging model based on two different supervised pipelines: a conventional feature-engineered Random Forest and a state-of-the-art end-to-end Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network. Based on the recordings of the Sleep-EDF Expanded database, we preprocessed the multichannel signals (EEG and EOG) into 30-second epochs, which were classified into five stages (Wake, N1, N2, N3 and REM). Our findings show that the feature-engineered baseline is very competitive with 95.49% accuracy, but the CNN-LSTM model performs better (96.44% accuracy) and has a much higher capability of classifying the transitional N1 stage which is harder. Because these results are obtained from a single overnight recording with a within-subject train/test split, they are best interpreted as a controlled proof-of-concept upper bound rather than as evidence of clinical-grade or human-level performance; cross-subject (leave-one-subject-out) validation on multiple recordings is required before any such claim can be made.