Comparative analysis of traditional machine learning Vs deep learning for sleep stage classification

Sleep stage classification is crucial for diagnosing sleep disorders and understanding sleep physiology. This study presents a comprehensive comparison between traditional machine learning algorithms and deep learning architectures using EEG recordings from the Physionet database. We extract 23 time and frequency domain features from each 30-second EEG segment and evaluate their performance across SVM, Random Forest, k-NN, and Gradient Boosting against CNN, LSTM, and hybrid CNN-LSTM models with attention mechanisms. Our results demonstrate that while traditional approaches achieve 82.4% accuracy with significant interpretability advantages, deep learning models reach 89.7% accuracy but require substantially more computational resources. The CNN-LSTM architecture with attention mechanisms performs best across all sleep stages, particularly for discriminating between similar stages like S1 and REM. However, traditional Random Forest classifiers offer competitive performance for resource-constrained applications with only 15% longer inference time. This comparative framework provides valuable insights for researchers and clinicians selecting appropriate methodologies for sleep analysis based on their specific requirements for accuracy, interpretability, and computational efficiency.