Advanced Machine Learning for classification of EEG traits as Parkinson’s biomarker

Abstract

We aim to develop non-­‐invasive, low-­‐cost preclinical markers for synucleinopathies (Parkinson’s Disease -­‐ PD or Dementia with Lewy Bodies -­‐ DLB) with impact on neuroprotection. In a first stage we have applied Machine Learning techniques to the analysis of the spontaneous, waking EEG of REM Sleep Behavior Disorder (RBD) patients. We know that patients with RBD may evolve towards PD and other synucleinopathies (i.e. DLB) providing a useful guide in the search for biomarkers. EEG has proven to be sensitive to brain alterations related to both RBD and PD patients, and it is known that patients with either RBD or de novo PD without obvious cognitive alterations show similar EEG/MEG alterations (mainly slowing of the EEG) while awake. In the work described herein we achieve the classification of patients according to their ultimate diagnosis, differentiating among different patient groups. We have analyzed EEG data from a study of RBD patients at the Center for Advanced Research in Sleep Medicine of Montreal. The EEG was recorded from 8 patients who years later evolved to PD, 8 patients who later evolved to LBD, 10 patients with RBD who did not convert, and 17 healthy controls. It is worth mentioning that 80% of the PD and DLB patients developed disease at a follow-­‐up of 8 years. Support Vector Machines were applied to different feature types. Absolute Band Power features extracted each 4 seconds outperform the other ones. After non-­‐parametric feature selection of the 5 most discriminative channel-­‐band combinations, classification was applied. A procedure for performance estimation close to operational conditions, which is denoted as leave-­‐pair-­‐subjects-­‐ out, has been employed for evaluation of the implemented system. Excellent performance levels, AUC 93.75-­‐99%, were obtained for all realized group comparisons. In conclusion, we believe that these results support the idea that classification of EEG shows great potential as a preclinical biomarker. Our results confirm the existent literature, where EEG slowing characterizes RBD and PD/DLB patients. Furthermore we have attained this characterization at the individual subject level for the first time by employing machine learning. Given that EEG was acquired 8 years before disease development in a majority of cases, we believe that the utilization of the developed system for conversion prognosis of synucleinopathies (PD or DLB) will have real impact on both early treatment and drug development.