Development and validation of artificial intelligence algorithms for Subthalamic Nucleus identification from intracortical signals.

Parkinson’s Disease (PD) is a neurodegenerative disorder that is rather common in our society nowadays which affects both movement and cognitive function depending on its stage. Since a definitive cure does not exist and the standard medication for PD are subjected to the issue of “wearing off”, one way to improve the quality of life of PD patients is the SubThalamic Nucleus Deep Brain Stimulation (STN-DBS). This operation consists of a surgical implantation of small electrodes in the STN to deliver electrical impulses which help to alleviate neurological and motor symptoms like tremors and rigidity. The effectiveness of STN-DNS is very much dependent on the accuracy of the electrode’s implantation. For this reason, Micro-Electrodes Recordings (MERs) are obtained during the operation to help identifying the position of the STN and ensuring the correct placement of the electrodes. In this work, two Artificial Intelligence (AI) algorithms are developed and validated to correctly identify the STN from the MERs and, possibly, to help clinicians during STN-DBS neurosurgery. The two explored algorithms are: Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models. To develop and validate these two models, a dataset of one-second MERs segments is obtained from a dataset of 36 PD patients that underwent bilateral STN-DBS, each segment is either labelled STN-IN or STN-OUT. After a prior process of hypertuning on a fraction of the total dataset, the CNN model trained with the complete dataset obtained an accuracy of 94.9% on the training set, of 73.0% on the validation set and of 73.8% on the test set. Afterwards, a k-fold cross-validation was performed on seven folds containing MERs segments each from different PD patients with a mean accuracy of 80.0% ± 2.2% across the seven folds. In addition, for each tested segment the computational time is approximately 37 ms, allowing the model to possibly be used in real-time applications. On the other hand, the LSTM model didn’t perform as well as the CNN model with a maximum accuracy of 79.9% during the hypertuning phase, required a much higher computing power. Since the results were still promising, it’s possible that in future works LSTM models for STN identifications could be explored further. Subsequently, a second dataset was obtained from a publicly available collection of MERs signal from 14 PD patients who also underwent STN-DBS to further validate the performances of the CNN and evaluate its generalization ability. On this dataset, the model performed noticeably worse, with an accuracy of 52.7%. This difference in performance can be explained by the different acquisition system used in the second dataset and by the fact that the segments are not standardized in amplitude but only in mean and standard deviation and that difference in amplitude may explain these different results.