Machine Learning-Based Polysomnography Data Analysis for ADHD Diagnosis: A Focus on Sleep Stage-Based Biomarkers

Attention-Deficit/Hyperactivity Disorder (ADHD) significantly impacts children and adults worldwide, posing challenges in neurodevelopmental diagnostics. This study aimed to improve ADHD diagnostic accuracy by identifying sleep stage-based biomarkers for ADHD through the analysis of polysomnography (PSG) recordings, including EEG, EOG, EMG, and ECG signals across five sleep stages. Our analysis included 20,294 epochs for ADHD and 19,968 epochs for non-ADHD from 48 patients. We applied graph metrics, particularly the average shortest path length, to quantify the efficiency of information transfer within brain networks and other physiological systems. We employed nested cross-validation to ensure robust model training and evaluation, with an outer 5-fold cross-validation and an inner GridSearchCV for hyperparameter tuning and model selection. The Random Forest model achieved an average accuracy of 72%, precision of 71%, recall of 85%, and F1-score of 76%. Feature Importance Analysis and Correlation Analysis highlighted that Sleep Stage 1 and Sleep Stage 3-4 were the most influential features for distinguishing ADHD from non-ADHD individuals. Permutation analysis further validated these findings. These results underscore the potential of sleep stage correlations and graph-based metrics as biomarkers for ADHD, demonstrating the model's ability to accurately classify ADHD using comprehensive PSG data. The implementation code has been made public https://github.com/ae1028/ML-PSG-ADHD-analysis