You are in:Home/Publications/A novel deep learning model for STN localization from LFPs in Parkinson's disease

Ass. Lect. Mohamed Hosny Ali Mohamed Elnogomy :: Publications:

Title:
A novel deep learning model for STN localization from LFPs in Parkinson's disease
Authors: Mohamed Hosny; Minwei Zhu; Wenpeng Gao; Yili Fu
Year: 2022
Keywords: Convolutional neural network; Genetic algorithm; Subthalamic nucleus localization Parkinson's; Parkinson's disease; Deep brain stimulation; Local field potentials
Journal: Biomedical Signal Processing and Control
Volume: 77
Issue: Not Available
Pages: 103830
Publisher: Elsevier
Local/International: International
Paper Link:
Full paper Mohamed Hosny _A novel deep learning model for STN localization from LFPs in Parkinson's disease.pdf
Supplementary materials Not Available
Abstract:

Deep brain stimulation (DBS) is a common treatment for the neurological disorder, Parkinson's disease (PD). DBS encompasses accurate implantation of stimulated electrodes at subthalamic nucleus (STN). Although imaging modalities are basically employed to identify STN boundaries, additional electrophysiological signals, microelectrode recording (MER), are important to deduce the underlying anatomy. Despite the adequate clinical outcomes, MER jeopardizes the patient safety. Recently, local field potentials (LFP) gained attention for STN detection due to their correlation with the motor territory of STN and ability to expose valuable insight into PD mechanisms. We propose a novel model, called CNN-GA-KNN, that improves the accuracy of STN localization using LFP. Convolutional neural network (CNN) model is used to extract deep local features. The generated features are fed to genetic algorithm (GA) technique for feature selection. The most distinguishing features are used as input of k-Nearest Neighbour (KNN) classifier, developed using leave-one-out strategy. Results show that KNN reached an average accuracy of 87.27%, outperforming the existing feature extraction techniques. CNN-GA- KNN betters five end-to-end deep learning approaches, yielding significant improvements in classification accuracy up to 7.47%, 8.48%, 7.06%, 6.93% and 6.86% in comparison to LSTM-CNN, GRU-CNN, BiLSTM-CNN, VGG16 and ResNet18, respectively. Results strongly imply that CNN-GA-KNN has tremendous potential to divulge informative features concerning STN taking advantage of a robust feature extraction strategy that combines the repeatability and learnability of GA and KNN. The system is ready to assist as an additional tool for fine-tuning the electrode trajectory in DBS centers instead of the current framework utilizing MER.

Google ScholarAcdemia.eduResearch GateLinkedinFacebookTwitterGoogle PlusYoutubeWordpressInstagramMendeleyZoteroEvernoteORCIDScopus