Explainable machine learning algorithm predicting working memory performance in Parkinson's disease using task-fMRI

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Parkinson's disease (PD) impairs motor and cognitive functions, especially working memory (WM).
• Task-based functional magnetic resonance imaging (fMRI) shows potential for decoding brain activity but has limited application in PD.
• Developing advanced analytical tools is crucial for understanding PD's impact on cognitive functions.

Purpose of the Study:

• To develop an explainable machine learning model for classifying WM performance levels in PD patients.
• To utilize task-based fMRI data for objective assessment of cognitive function in PD.
• To enhance the interpretability of neuroimaging findings in Parkinson's disease research.

Main Methods:

• 45 PD patients and 15 healthy controls (HCs) underwent task-based fMRI during an n-back WM task.
• PD patients were stratified into better, intermediate, and worse WM performance subgroups based on 3-back task results.
• A 3D convolutional neural network (3D-CNN) model, pre-trained with a 3D autoencoder, was employed for binary classifications.

Main Results:

• The 3D-CNN model achieved 93.3% accuracy in distinguishing PD patients with worse WM from HCs.
• This accuracy significantly surpassed the mean accuracy of expert radiologists (70.0%).
• Saliency maps highlighted the dorsolateral prefrontal cortex and parietal lobules as critical regions for WM performance, consistent with fMRI data.

Conclusions:

• An explainable deep learning model was successfully developed to classify WM performance in PD using task-based fMRI.
• This approach offers an objective and interpretable method for assessing brain function in clinical neuroimaging.
• The findings suggest potential for improved diagnostic and prognostic tools in Parkinson's disease management.