Data-driven classification algorithms for identification and characterization of early neurodegeneration

Parkinson’s disease (PD) is a common neurodegenerative disease and is diagnosed in presence of motor symptoms, which appear when the neurodegeneration has affected a large part of the brain. No treatment is currently available to slow down or stop the neurodegeneration. Rapid eye movement (REM) sleep behavior disorder (RBD) is a sleep disorder characterized by abnormal increase in muscular activity during REM sleep and dream enactment.

RBD is the strongest early biomarker of PD. Evidence show that RBD might be preceded by a prodromal phase, where minor REM behavioral events can be appreciated but not enough abnormal muscular tone is seen to diagnose RBD. Currently, identification of RBD and prodromal RBD is based on visual inspection of video, electroencephalography (EEG), electrooculography (EOG) and electromyography (EMG) recorded during sleep.

Such a process is slow and not objective. This thesis, divided into three parts, presents new data-driven automated methods for fast and objective identification and characterization of RBD and prodromal RBD. First, a comparison of the currently available automated methods for RBD detection was performed.

The results showed that none of them could be considered the optimal one, due to varying performances when they were applied to different patient groups. Second, a new data-driven method for RBD identification, based on machine learning techniques applied to EMG signals, was developed. The method showed higher performances for identifying RBD patients than previously developed methods (accuracy and specificity over 80% and sensitivity over 70%).

Moreover, it was found that muscular activity in non-REM sleep contributed for more accurate RBD identification. When the new method was applied to data recorded in another clinic, it identified RBD patients with similar performances, thus proving its robustness. Finally, a data-driven method based on machine learning applied to EEG and EOG signals was developed to identify RBD and prodromal RBD in PD patients.

RBD was identified with over 80% accuracy, sensitivity and specificity. Moreover, the algorithm could identify PD patients with prodromal RBD with accuracy and specificity over 80%. It was also found that micro-sleep instability could be a biomarker for RBD and prodromal RBD in PD patients. These findings are potentially applicable to patients without overt PD.

In conclusion, this thesis proposes new, fast and automatic methods and biomarkers to identify patients with RBD and prodromal RBD. Compared to current visual-based methods, these algorithms have the potential to be used to identify patients in early stages of neurodegeneration objectively, consistently and significantly faster.

Thanks to the objective identification, such patients could constitute a homogeneous target for future neuroprotective trials aiming at slowing down or even stopping the ongoing neurodegeneration.