Identification of Depression Subtypes in Parkinson's Disease Patients via Structural MRI Whole-Brain Radiomics: An Unsupervised Machine Learning Study

Abstract

Objective

Current clinical evaluation may tend to lack precision in detecting depression in Parkinson's disease (DPD). Radiomics features have gradually shown potential as auxiliary diagnostic tools in identifying and distinguishing different subtypes of Parkinson's disease (PD), and a radiomic approach that combines unsupervised machine learning has the potential to identify DPD.

Methods

Analyze the clinical and imaging data of 272 Parkinson's disease (PD) patients from the PPMI dataset, along with 45 PD patients from the NACC dataset. Extract radiomic features from T1-weighted MRI images and employ principal component analysis (PCA) for dimensionality reduction. Subsequently, apply four unsupervised clustering methods including Gaussian mixture model (GMM), hierarchical clustering, K-means, and partitioning around medoids (PAM) to classify cases in the PPMI dataset into distinct subtypes. Identify high-risk subtypes of DPD on the basis of the time and number of depression progression, and validate these findings using the NACC dataset. The data from the high-risk subtype were divided into a training subtype and a testing subtype in a 7:3 ratio. Multiple logistic regression analysis was conducted on the training subtype data to develop a traditional logistic regression model for the high-risk subtype, which was subsequently compared with a supervised logistic regression model constructed for the entire PPMI cohort. Finally, the performance of both models was evaluated using receiver operating characteristic (ROC) curves. In addition, a decision tree (DT) model was constructed based on independent risk factors of high-risk subtypes and validated using low-risk subtype data. ROC curves were employed to validate this model across training subtype, testing subtype, and low-risk subtype datasets.

Results

The PAM clustering method demonstrates superior performance compared to the other three clustering methods when the number of clusters is 2. High-risk subtypes of DPD can be effectively distinguished in both the PPMI and NACC datasets. A traditional logistic regression model was developed based on rapid-eye-movement behavior disorder, UPDRS I score, UPDRS II score, and ptau in high-risk subgroups. This model exhibits a diagnostic efficacy (AUC = 0.731) that surpasses that of the traditional regression model constructed using the entire PPMI cohort (AUC = 0.674). The prediction model based on high-risk subtypes had AUC values of 0.853 and 0.81 in the training and testing subtypes, sensitivities of 0.765 and 0.786, and specificities of 0.771 and 0.815, respectively. The AUC, sensitivity, and specificity in the nonhigh-risk subtype were 0.859, 0.654, and 0.852, respectively.

Conclusion

By identifying MRI structural radiomics and clinical features as potential biomarkers, the radiomic approach and UCA provide new insights into the pathophysiology of DPD to support the clinical diagnosis with high prediction accuracy.