Advancing narcolepsy diagnosis: Leveraging machine learning to identify novel neuro-biomarkers

Abstract

Narcolepsy is a rare neurological disorder with a well-identified pathophysiology that manifests as a sudden onset of sleep during wake behaviour. The current diagnostic pathways for narcolepsy involve complex assessments of sleep neurophysiology, including polysomnography and the multiple sleep latency (MSLT) test. These are cumbersome and work-intensive, and with limited resources within the NHS, this has led to increased waiting times for diagnosis and treatment of narcolepsy. This project harnessed the power of digital neuro-biomarkers and Artificial Intelligence (AI) to develop novel diagnostic markers for narcolepsy. Leveraging an open-source dataset of labelled archival polysomnography (PSG) recordings, including electroencephalography (EEG), I created a data analysis and classification pipeline to enhance diagnostic decision-making in clinical settings. This pipeline combines comprehensive data preprocessing and feature extraction with XGBoost and Random Forest (RF) classification models. The feature extraction process included selected time- series analysis features, spectral frequency ratios, cross-frequency coupling and moment-based statistical features of Intrinsic Mode Functions (IMFs) derived from empirical mode decomposition (EMD). The RF classifier emerged as the best model, achieving an accuracy of 82.5%, with a specificity of 82.5% and a sensitivity of 92.86%, by combining and averaging these feature sets and incorporating sleep stage labels during model training. These results underscore the potential of a novel approach using single-channel sleep EEG data from wearable devices. This innovative method simplifies the lengthy and costly pathway for narcolepsy diagnosis and also paves the way for developing new tools to diagnose sleep disorders automatically in non-clinical environments.