Objective: To test whether machine learning algorithms are suitable to classify dopamine and behavioral states in Parkinson’s patients (PD).

Background: While motor symptoms of PD patients can be effectively treated by continuous deep brain stimulation (DBS), an adaptive or closed loop stimulation (CLS) may grant many benefits, including the possible reduction of side-effects. For this purpose, biomarkers are needed which reliably detect the patient’s current pathological and/or behavioral state.

Methods: After implantation of DBS electrodes in the subthalamic nucleus, three PD patients (51-63 years) performed a rest and a tapping task using their right index finger. During the task, local field potentials were recorded from the implanted electrodes at a sampling rate of 2500 Hz. Both tasks were performed in the medical ON and the OFF state. Using accelerometric data, single taps were epoched into smaller windows of 0.5 s length. Similarly, rest trials were epoched into windows of the same size. From each trial, mean normalized power was extracted in the theta (4-8 Hz), beta (13-30 Hz) and gamma-band (30-100 Hz) using a prolate spheroidal sequences based multitaper. We applied three different machine learning approaches to 1) classify the behavioral state of the patients (i.e. rest vs. tapping) and 2) to classify the dopamine-state (i.e. ON vs OFF). A k-fold cross validation (k=10) was used to avoid overfitting.

Results: ON vs. OFF: In each of the three patients we archieved an accuracy of >96% in the tapping condition and >=97% in the rest condition. Rest vs. Tapping: In each of the three patients we achieved an accuracy of >66% in the ON-state and >76 % in the OFF-state.

Conclusions: Machine learning is highly suited to classify the dopaminergic state of PD patients. Wether the applied algorithms are usable in an online classification scheme suitable for CLS needs to be further evaluated.

« Back to 2018 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/classification-of-dopaminergic-state-and-motor-activity-by-theta-beta-and-gamma-activity-in-parkinsons-disease/