The use of Machine Learning models in the diagnosis of Parkinson’s Disease with fMRI data: a Systematic Review and Meta-Analysis

Background and aims: Parkinson’s Disease (PD) is a commonly occurring neurodegenerativedisorder worldwide, affecting approximately ۱% of individuals over the age of ۶۰. Early detectionof PD is crucial in managing symptoms and improving the quality of life for patients. Resting-state functional magnetic resonance imaging (rs-fMRI) has shown to be widely accepted fordetecting inherent brain function changes during the early stages, enabling early diagnosis beforestructural changes emerge. The use of Artificial Intelligence (AI) has proven successful in diagnosingother conditions, such as breast cancer. Therefore, we can expect more accurate detectionof PD using AI to improve diagnostic outcomes.Method: A comprehensive systematic search using relevant keywords such as “fMRI”, “ArtificialIntelligence”, “machine learning”, and “Parkinson’s disease” was conducted on four MajorOnline Databases; PubMed, Scopus, Web of Science, and Embase up to March ۲۰۲۳. The searchof the database encompassed not only published literature but also grey literature as well as manualsearch. After the initial search, two independent reviewers screened the retrieved publications,ensuring that all studies that followed the criterion of employing Artificial Intelligence (AI) modelsor machine learning algorithms for predicting or diagnosing Parkinson’s Disease with the useof fMRI data were included. Subsequently, the studies that met the inclusion criteria underwenta critical appraisal by two authors independently. The quality of studies was evaluated accordingto the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-۲) checklist. Furthermore,for data extraction purposes, Microsoft Office Excel ۲۰۲۱ software was employed to collect informationsuch as machine learning algorithms, the accuracy of models, and fMRI Image properties.Pooled Accuracy for detected PD was calculated using CMA v.۳.۷ software and a p-value lessthan ۰.۰۵ was considered a significant level.Results: We retrieved ۱۱۲۲ relevant studies from online databases. After a thorough examinationof the titles and abstracts and the removal of duplicate publications(n=۲۷۷), ۷۱۰ studies wereeliminated. In ۲۵ cases of disagreement between two authors, the opinion of the third author wasthe determiner. The full texts of ۱۴۰ papers were reviewed. Eventually, ۱۳ studies met our inclusioncriteria and were included in this study. All of the studies have different levels of bias. However,regarding the type of study design of included articles are considered as low risk. Approximately۷۷% of studies utilized the SVM (Support Vector Machine) algorithm for differentiatingpatients with Parkinson’s disease from healthy controls. The overall accuracy was determined tobe ۸۳.۸% using the random effects model (accuracy = ۰.۸۳۸, ۹۵% CI = ۰.۷۸۸-۰.۸۷۸, p-value <۰.۰۰۱), indicating a significant predictive power.Conclusion: Machine Learning can be a useful tool in early PD diagnosis and could also aid inprognosis orientation at lower costs. Nevertheless, the possibility of overfitting in machine learningalgorithms means that more research is necessary to determine the full potential of this approach.While machine learning can be a valuable tool in early PD diagnosis, the limited numberof cases studied in current literature must be taken into consideration.Additionally, since fMRI data contains a vast amount of information, overfitting remains a significantconcern. However, the SVM algorithm’s widespread use in these studies is due to itsresilience to overfitting.

Keywords:

Parkinson’s Disease (PD) , Machine Learning (ML) , Functional Magnetic Resonance Imaging(fMRI) , Diagnosis