Real-Time Processing and Classification of Intracortical Recordings during STN-DBS Neurosurgery

Deep Brain Stimulation of the SubThalamic Nucleus (STN-DBS) has been proven successful in alleviating symptoms and improving the quality of life in patients suffering from Parkinson’s Disease (PD). It involves the implantation of a stimulating electrode within the STN and its accurate placement is essential for achieving optimal therapeutic outcomes. To this aim, MicroElectrode Recordings (MERs) are acquired during the surgery to provide real-time visual and auditory confirmation of the electrode position. The objectives of this thesis focus on improving the accuracy of STN targeting and gaining deeper insight into MER properties. For these purposes, a dataset of MERs from 36 PD patients, who underwent bilateral DBS electrode implantation, is retrospectively analyzed. For each patient, while one hemisphere was always kept in control condition, the other was operated in one of the following experimental conditions: wearing headphones and listening to nothing, listening to music, or listening to white noise. MERs are often corrupted by motion or electromagnetic artifacts. Therefore, this thesis introduces a novel artifact detection and removal algorithm that compares variances of adjacent signal windows. The algorithm parameters are optimally chosen to minimize an objective function that balances the trade-off between artifact removal rate and STN signal preservation. Then, 31 features in the time and frequency domain are extracted, and Feature-Feature Correlation analysis is conducted to identify potential redundancies among them. Different machine learning algorithms, including k Nearest Neighbors (kNN), Decision Tree (DT), Support Vector Machine (SVM), and MultiLayer Perceptron (MLP), are trained with different feature combinations and validated using leave-one-patient-out cross-validation. The selection of features’ subsets is driven by the feature-ranking methods Maximum Relevance Minimum Redundancy (MRMR) and ReliefF, and an algorithm based on scatter matrices. The classifier that offers the best trade-off between real-time applicability and accuracy is integrated into a user-friendly Graphical User Interface (GUI), designed in MATLAB® to support neurosurgeons during the decision-making process. As a final objective, this research investigates the influence of hemisphere (right or left) and auditory condition on MER characteristics conducting an ANOVA on the extracted features. Feature-Feature Correlation analysis reduced the number of key features to 11. The classifier chosen to be integrated in the GUI is the MLP that takes as inputs the first 3 features from the MRMR ranking algorithm. It achieves an average accuracy of 86.4 ± 5.9% (mean ± standard error), with processing time for feature extraction and prediction shorter than 20ms. The average accuracy increases to 91.4 ± 1.6%, when the Estimated Distance from Target (EDT) is employed as an additional feature. Results indicate that few informative features can be used for an efficient classification of MERs to be employed intraoperatively to support clinical decision. ANOVA revealed some statistically significant differences between hemispheres when analyzing features like PSD index, that is an indicator of the signal power in the band [250Hz, 2.5kHz], and conditions when analyzing attributes like RMS. These preliminary findings pave the way for future in-depth investigations that could unveil more relevant properties.