Autoencoder-based Feature Extraction and Classification for fMRI-based Deep Brain Stimulation Parameter Optimization for Parkinson's Disease Treatment: Towards a Rapid Semi-automated Stimulation Optimization

Optimized deep brain stimulation (DBS) is fast becoming a therapy of choice for the treatment of Parkinsons disease (PD). However, the post-operative optimization (patient clinical benefits are maximized and adverse effects are minimized) of the large number of possible DBS parameter settings (signal frequency, voltage, pulse width and contact locations) using the current empirical protocol requires numerous clinical visits, which substantially increases the time to reach optimal DBS stimulation, patient cost burden and ultimately limits the number of patients who can undergo DBS treatment. These issues became even more problematic with the recent introduction of electrode models with stimulation directionality thereby enabling more complex stimulation paradigms. These difficulties have necessitated the search for a biomarker-based optimization method that will streamline the DBS optimization process. Our recently published functional magnetic resonance imaging (fMRI) and machine learning-assisted DBS parameter optimization for PD treatment has provided a way to rapidly classify DBS parameters using parcel-based features that were extracted from DBS-fMRI response maps. However, the parcel-based method had limited accuracy as the parcels are based on subjective literature review. Here, we propose an unsupervised autoencoder (AE) based extraction of features from the DBS-fMRI responses to improve this accuracy. We demonstrate the usage of the extracted features in classification methods such as multilayer perceptron (MLP), random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN) and LDA. We trained and tested these five classification algorithms using 122 fMRI response maps of 39 PD patients with a priori clinically optimized DBS parameters. Further, we investigated the robustness of the AE-based feature extraction method to changes in the activation patterns of the DBS-fMRI responses, which may be caused by difference in stimulation side and disease condition. Changes in the locations of activated and deactivated brain regions was simulated using a left-right horizontal flipping of the original left-sided (or nominal) DBS-fMRI response maps. The visualization of AE-based features extracted from the nominal and flipped DBS-fMRI response maps formed optimal and non-optimal clusters in a neuro-functionally meaningful manner, which indicate robustness of the AE-based feature extraction to subtle differences in the activated regions of DBS-fMRI response maps. The MLP, RF, SVM and LDA methods gave an overall DBS parameter classification accuracy of 96%, 94%, 92% and 93% respectively when trained using the AE-extracted features from the nominal DBS-fMRI maps. The AE-based MLP, RF, SVM and LDA accuracies were higher than the overall accuracy (81%) of our initial parcel-based LDA method. The performance of an AE-MLP model trained using the nominal DBS-fMRI maps did not change significantly when the model was tested on the flipped DBS-fMRI responses. We showed that the MLP method combined with AE-based feature extraction is best suited for fMRI-based DBS parameter optimization and represents another step towards a proposed digital tool for rapid semi-automated biomarker-based DBS optimization.