Processing of intraoperative recordings during Subthalamic Nucleus Deep Brain Stimulation Neurosurgery in patients with Parkinson Disease

Parkinson’s disease is a chronic and progressive nervous system disorder that affects movement. To improve the quality of life of patients, medication and treatments have been developed, among which SubThalamic Nucleus Deep Brain Stimulation (STN-DBS) is thought to be the most effective treatment. STN-DBS consists of the surgical implantation of microelectrodes in the basal ganglia and the delivery of an electric stimulation. However, the effectiveness of STN-DBS is dependent on the accuracy of the implantation. For this purpose, microelectrode recordings are obtained during the surgery to better identify the target location. In this work, we devise a simple yet effective machine learning method that, given microelectrode recordings, enables the detection of the SubThalamic Nucleus with high accuracy and explainability. Using decision trees, we achieve an accuracy of 92.1% and F1-score of 90.0% on a dataset of 36 patients. While this method is known to lack robustness and stability due to the direct dependence on the training set, we compensate for these limitations by employing decision tree ensembles, boosting the model stability as well as the observed performance significantly to 94.3% accuracy and 92.7% F1-score. Our algorithm is therefore a step forward to a reliable prediction of the STN for improved treatment, reducing the dependence on the specialist. The second part of this work focuses on music therapy which is a noninvasive treatment of Parkinson’s disease. Music therapy uses auditory stimulation of the patient and has shown to be effective in improving motor symptoms, such as gait parameters. While the physiological background of music therapy is not yet well understood, we aim in this work at shedding light on understanding the effects of music therapy on the subthalamic nucleus neuronal activity of Parkinson’s disease patients. Therefore, we analyze microelectrode recordings of brain hemispheres under different auditory stimulation conditions. More precisely, one hemisphere was always in control condition (silence during the surgery), while the other hemisphere was either in basal or experimental condition (music or noise was played during the signal acquisition time). To investigate if the different conditions can be automatically identified, we extracted features from the recordings and applied clustering. Our preliminary results indicate that representing the patients by the difference between the features of their both hemispheres, we can successfully separate the patients from the experimental group from those of the control group. Inside the experimental group, however, we did not find a separation between the music and the noise condition. We hypothesize that the impossibility to separate these groups could be due to the high inter-subject and inter-hemisphere variability.