نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Comparison of Static, Dynamic, and Combined Functional Connectivity Features Across Different Brain Parcellations for Parkinson’s Disease Classification Using Machine Learning

Subject:

Artificial Intelligence and Brain Mapping

Presenter Author:
Name:	Mostafa Salehi

Author:
Name 1:	Mostafa Salehi - Corresponding - Presenter
Affiliation 1:	College of Interdisciplinary Science and Technologies, University of Tehran

Name 2:	Maryam Soltani
Affiliation 2:	College of Interdisciplinary Science and Technologies, University of Tehran

Name 3:	Reza Nazari
Affiliation 3:	College of Interdisciplinary Science and Technologies, University of Tehran

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Parkinson's disease (PD) is clinically characterized by its motor signs, yet its functional impact extends well beyond motor circuits. Resting-state functional magnetic resonance imaging (rs-fMRI) captures these functional alterations by measuring temporal coupling among brain regions. Although rs-fMRI features can separate individuals with PD from healthy controls, reported performance varies widely, largely because analytic pipelines differ—most notably in the choice of brain parcellation. We therefore compared static, dynamic, and combined functional connectivity (FC) features across five parcellations and four classifiers to determine how each component shapes classification performance.

Methods:

Data and preprocessing. Structural T1-weighted and resting-state blood-oxygen-level-dependent (BOLD) images were obtained from the Parkinson's Progression Markers Initiative (PPMI); after quality review, 164 participants (112 patients with PD, 52 healthy controls) were included. Preprocessing used fMRIPrep 25.2.3, with head-motion and slice-timing correction, BOLD-to-T1w registration, and nonlinear normalization to MNI152NLin2009cAsym. Because scan durations differed, the first 210 time points were retained for every participant. Feature extraction and classification. Five parcellation schemes defined the network nodes: AAL116, Harvard-Oxford48, Schaefer100, K-means100, and Ward100. For each parcellation, static FC was computed as Pearson correlations between regional time series, and dynamic FC was derived from a sliding-window analysis (window length: 30 time points, step: 15), yielding 17 summary features that captured temporal variability in connectivity; combined features concatenated the static and dynamic vectors. Four classifiers were evaluated: logistic regression, linear support vector machine (SVM), random forest, and XGBoost, with SelectKBest (ANOVA F) for feature selection. A nested cross-validation design (5 outer folds, 3 inner folds) governed training and evaluation, with every data-dependent step—imputation, standardization, feature selection, and hyperparameter tuning—confined to the training folds to prevent leakage. In total, 60 configurations were assessed.

Results:

Best configuration. The strongest configuration—AAL116 with static FC and random forest—achieved a receiver operating characteristic area under the curve (ROC-AUC) of 0.8934, with accuracy 0.8902, balanced accuracy 0.8630, sensitivity 0.9375, specificity 0.7885, precision-recall area under the curve (PR-AUC) of 0.9332, and F1-score 0.9211. Comparison across parcellations. AAL116 showed the highest average ROC-AUC, but its advantage over Schaefer100 and Harvard-Oxford48 was not statistically significant, whereas K-means100 and Ward100 averaged substantially lower (0.5945 and 0.6308, respectively). Comparison across feature types. Across all configurations, combined features achieved the highest average ROC-AUC (0.7631), followed by dynamic (0.7237) and static (0.7154) features, although the single best model relied on static features alone. Comparison across classifiers. The four algorithms performed comparably on average (ROC-AUC range: 0.7255–0.7402), and pairwise comparisons revealed no significant differences after correction for multiple testing.

Conclusion:

Classification performance varied across combinations of parcellation, feature type, and classifier rather than tracking any single component. The choice of parcellation and feature type, in other words, may matter as much as the classifier itself in PD classification. Because these results rest on internal validation, and because PD is clinically heterogeneous, they should not be assumed to transfer to other cohorts.

Keywords:

Parkinson's Disease, Functional Connectivity, Resting-State fMRI, Machine Learning, Brain Parcellation, Classification