Deriving mechanistic insights from machine learning and its possible implications in non-invasive brain stimulation research

Classification of Parkinson's disease motor phenotype: a machine learning approach

To assess the cortical activity in people with Parkinson's disease (PwP) with different motor phenotype (tremor-dominant-TD and postural instability and gait difficulty-PIGD) and to compare with controls. Twenty-four PwP (during OFF and ON medication) and twelve age-/sex-/handedness-matched healthy controls underwent electrophysiological assessment of spectral ratio analysis through electroencephalography (EEG) at resting state and during the hand movement. We performed a machine learning method with 35 attributes extracted from EEG. To verify the efficiency of the proposed phenotype-based EEG classification the random forest and random tree were tested (performed 30 times, using a tenfolds cross validation in Weka environment). The analyses based on phenotypes indicated a slowing down of cortical activity during OFF medication state in PwP. PD with TD phenotype presented this characteristic at resting and the individuals with PIGD presented during the hand movement. During the ON state, there is no difference between phenotypes at resting nor during the hand movement. PD phenotypes may influence spectral activity measured by EEG. Random forest machine learning provides a slightly more accurate, sensible and specific approach to distinguish different PD phenotypes. The phenotype of PD might be a clinical characteristic that could influence cortical activity.