To systematically investigate the robustness, generalization, interpretability, and privacy-preserving potential of voice-based Parkinson’s disease detection models under realistic deployment conditions.
Approach:
Research Questions: The study addresses eight research questions focusing on classification accuracy, dataset heterogeneity, acoustic feature robustness, explainable AI interpretations, federated learning performance, dataset aggregation effects, statistical significance of performance differences, and evaluation criteria beyond accuracy.
Key Findings:
High classification accuracy from single-dataset evaluations may not translate to robust performance in cross-dataset scenarios.
Dataset heterogeneity significantly affects the generalization capability of voice-based PD detection models.
Certain acoustic features, such as frequency-based (jitter) and amplitude-based (shimmer), may have varying contributions to PD detection.
Explainable AI techniques may provide stable interpretations under cross-dataset conditions.
Federated learning can achieve competitive generalization performance while preserving data privacy.
Interpretation:
The study highlights the need for robust evaluation methods that consider generalization and interpretability in the development of voice-based PD detection systems.
Limitations:
The study may be limited by the specific datasets used and their inherent biases, which could affect the generalizability of the findings.
The feature harmonization strategy may reduce the dimensionality of data, potentially impacting model performance.
Conclusion:
The proposed framework aims to enhance the clinical applicability of voice-based PD detection models by focusing on generalization and interpretability.
