Deep Transfer Learning for Parkinson’s Disease Monitoring by Image-Based Representation of Resting-State EEG Using Directional Connectivity

Abstract

Parkinson’s disease (PD) is characterized by abnormal brain oscillations that can change rapidly. Tracking neural alternations with high temporal resolution electrophysiological monitoring methods such as EEG can lead to valuable information about alterations observed in PD. Concomitantly, there have been advances in the high-accuracy performance of deep neural networks (DNNs) using few-patient data. In this study, we propose a method to transform resting-state EEG data into a deep latent space to classify PD subjects from healthy cases. We first used a general orthogonalized directed coherence (gOPDC) method to compute directional connectivity (DC) between all pairwise EEG channels in four frequency bands (Theta, Alpha, Beta, and Gamma) and then converted the DC maps into 2D images. We then used the VGG-16 architecture (trained on the ImageNet dataset) as our pre-trained model, enlisted weights of convolutional layers as initial weights, and fine-tuned all layer weights with our data. After training, the classification achieved 99.62% accuracy, 100% precision, 99.17% recall, 0.9958 F1 score, and 0.9958 AUC averaged for 10 random repetitions of training/evaluating on the proposed deep transfer learning (DTL) network. Using the latent features learned by the network and employing LASSO regression, we found that latent features (as opposed to the raw DC values) were significantly correlated with five clinical indices routinely measured: left and right finger tapping, left and right tremor, and body bradykinesia. Our results demonstrate the power of transfer learning and latent space derivation for the development of oscillatory biomarkers in PD.

Classifying Parkinson’s Disease Using Resting State Electroencephalogram Signals and U EN -PDNet

“…Utilizing the orthogonalized directed coherence (gOPDC) enumerating the directional connectivity (DC) between the pairwise EEG channels among the quartet frequency (θ,α,β, and γ) and visualizing the calculated DCs into 2D plots as the input to the selected network of VGG-16 (as the pre-trained model). It achieved the accuracy of 99.62% with precision 100% and recall of 99.17% along with 0.9958 of F1-score, and the AUC averaged over 10 repetitive instances of training proffers 0.9958 [25].…”

Ultra-Wideband Radar for Simultaneous and Unobtrusive Monitoring of Respiratory and Heart Rates in Early Childhood: A Deep Transfer Learning Approach

“…In our study, we utilized the VGG-16 architecture [ 23 ], which was pre-trained on the ImageNet dataset, including 1.2 million images. We only used the weights of convolutional layers from the pre-trained model.…”

“…It has been demonstrated that the primary layers of a Deep Convolutional Neural Network (DCNN) trained on a vast data set can extract generic features [ 20 , 21 ]. One favorable network in literature for transfer learning is the VGG-16 architecture [ 22 ] trained on the ImageNet dataset pre-trained model [ 23 ]. It has already been shown that the VGG-16 network is able to extract informative features for vital signs (HR/RR) estimation among adults [ 24 ].…”

A Comprehensive Review on AI-Enabled Models for Parkinson’s Disease Diagnosis

“…Transfer learning, which also provided a latent space that is more strongly related to clinical indices, produced excellent classification performance. According to the research of [185], a heatmap-integrated image classifier combined with transfer learning may provide a way to classify small sample datasets. The effectiveness of the classifier and research methodology will be further improved.…”