Generalization over accuracy: A cross-dataset, explainable, and federated learning framework for Parkinson's disease detection.

Objective: Parkinson's disease (PD) is a progressive neurodegenerative disorder for which early screening remains challenging. Although voice-based machine learning approaches have shown promise as non-invasive screening tools, most existing studies rely on single-dataset evaluations, random data splits, and accuracy-centric metrics, raising concerns about dataset bias, subject leakage, and limited real-world generalizability. This study aims to develop a generalization-aware, explainable, and privacy-preserving framework for PD detection using voice data. Methods: Three heterogeneous PD voice datasets are integrated using a strict feature harmonization strategy that retains only common acoustic features to enable fair cross-dataset evaluation. A diverse set of classical, ensemble, neural, and meta-learning models is evaluated under subject-aware experimental protocols with extensive cross-validation and hyperparameter optimization. The framework further incorporates ablation analysis and statistical significance testing. Realistic deployment conditions are simulated through cross-dataset generalization experiments, complemented by explainable AI (XAI) analysis using SHAP and LIME, as well as federated learning simulations for privacy-preserving training. Results: Experimental findings show that models achieving over 95% accuracy in single-dataset settings experience substantial performance degradation under cross-dataset evaluation, indicating strong dataset dependency in prior approaches. Performance typically converges to moderate levels, reflecting a generalization ceiling under dataset heterogeneity. Frequency-based perturbation features (jitter-related measures) consistently demonstrate greater robustness than amplitude-based features across datasets. Explainability analysis confirms the stability and physiological relevance of key acoustic biomarkers despite reduced predictive performance. Federated learning models achieve comparable or improved generalization performance relative to centralized training while preserving data privacy. Conclusion: By reframing voice-based PD detection as a generalization and trustworthiness problem rather than an accuracy optimization task, this study provides a more realistic and deployment-oriented evaluation framework. The findings highlight the importance of cross-dataset validation, robust feature selection, explainability consistency, and privacy-aware learning, offering a more clinically meaningful foundation for future healthcare AI systems.