Machine Learning Strategies for Single-Channel EEG Automatic Detection of REM Sleep Behavior Disorder: a Model Based on REM and Slow Wave Sleep

Human Sleep is the cyclic repetition of states characterized by different processes which play an important role in a wide range of activities, such as restoring the body's energy as well as supporting memory consolidation, and clearance of metabolic waste products generated by awake brain neural activity. It is mainly divided into two macro-stages, namely Rapid-Eye movement (REM) sleep and non-Rapid-Eye movement (NREM) sleep which, accordingly to the American Academy of Sleep Medicine (AASM) guidelines, is further characterized by three stages N1, N2, and N3 (or Slow Wave Sleep). In recent years, Sleep Disorders got the researchers' attention and some studies demonstrated a strong correlation with some types of Neurodegenerative Disease pathogenesis. Remarkably, Idiopathic REM Sleep Behavior Disorder (iRBD) shows the strongest correlation with the family of alpha-synucleinopathies, e.g. Parkinson's Disease (PD), Dementia with Lewy Bodies (DLB), and Multiple System Atrophy (MSA), and it is supposed to represent an early symptom of these neurodegenerative conditions with particularly high rate of phenoconversion into PD (up to 80% after 14 years). Recent studies also highlighted the potential role of disturbed Slow Wave Sleep (SWS) as a predictive biomarker of neurodegenerative processes that involve both PD and Dementia. This work aims to overcome the limits of the to-date available diagnostic tools proposing a fully-automatic EEG-based strategy able to detect RBD exploiting segments recorded during both REM and SWS sleep. Supervised Machine Learning models were trained and tested in a Leave-One-Out cross-validation framework, obtaining values of accuracy up to 91% and of sensitivity up to 94% (RBD class) and hence highlighting REM and SWS microstructures capabilities as RBD biomarkers. These results point to the potential of an EEG-based, low-cost, automatic RBD detection system which can be used in early diagnosis of neurodegeneration in order to spot prone individuals, and allowing them to join clinical trials of neuroprotective therapies to halt or at least delay the progression. The last part of the work concerns the development of a three-stages semi-supervised-based method to qualify REM Sleep without Atonia (RSWA) as an intermediate pathological state supporting a finer characterization of the neurodegenerative progression from Healthy to RBD. The achieved values of Rand-Index (97%) and Clustering Purity (99%) confirm the existence of peculiar RSWA EEG-patterns as well as suggest the reliability of the process as a basis from which to carry out deeper analysis.