Combining plasma biomarkers, clinical parameters, and neuroimaging features for differential diagnosis of Parkinson’s disease and atypical parkinsonian syndromes: a multidimensional modeling approach

Background: The early differential diagnosis of Parkinson's disease (PD) and atypical parkinsonian syndromes (APSs) poses challenges. The current methods, which rely on clinical assessments and single-modal biomarkers, lack sufficient sensitivity and specificity. This study aims to develop a multidimensional model integrating plasma biomarkers, clinical parameters, and neuroimaging radiomic features to improve the accuracy of differentiating PD from APS. Methods: -syn, and tau), clinical indicators (e.g., disease duration and UPDRS-III scores), and radiomic features (1,316 IBSI-standardized features) from magnetic resonance imaging scans of the midbrain and pons were collected. Core variables were screened using LASSO regression and random forest algorithms, and a multivariate logistic regression model was constructed. The performance of the model was evaluated using receiver operating characteristic curves, calibration curves, and cross-validation. Results: < 0.001), but there was no difference between the two groups. Four core variables (NFL, GFAP, disease duration, and pontine voxel volume) were selected. The area under the curve (AUC) of the combined model for identifying PD and APS was 0.874 (95% confidence interval: 0.801-0.946), which was significantly higher than that of single variables (such as NFL alone AUC = 0.653). Cross-validation confirmed the stability of the model (AUC = 0.843). Conclusion: This was the first study to integrate plasma NFL/GFAP, clinical disease duration, and pontine radiomic features to construct a high-precision PD-APS differential model (AUC > 0.87), addressing the limitations of traditional single-mode approaches. The gradient changes in GFAP and the APS-specificity of NFL serve as key biomarkers. Thus, this multidimensional framework provides a practical diagnostic tool for resource-limited scenarios.