Parkinson’s Disease Detection Using Dual-Branch WavLM Embeddings, Fuzzy Cellular Automata Statistics, and Mamba Sequence Modeling

Abstract

Parkinson′ s disease(PD) is a progressive neurodegenerative condition that can affect the muscular control of speech at an early stage. This work presents an FCA–Mamba framework for automatic Parkinson’s disease detection from speech using contextual representation learning and uncertainty-aware spatial modeling. Preprocessed speech waveforms are processed by a pretrained, frozen WavLM encoder to obtain contextualized speech embeddings. The Mamba branch can selectively model long-distance temporal dependencies in contextual speech representations, while the FCA branch exploits fuzzy neighborhood evolution and standalone higher-order statistics for localized uncertainty-aware spatial relationships. These representations are then fused temporally and statistically using the two branches to produce final Parkinson speech classification. Experiments are conducted using stratified five-fold cross-validation on an established dataset of 655 speech samples from people with Parkinson’s disease. The proposed FCA–Mamba framework achieves 98.68% accuracy, 98.58% precision, 99.49% recall, 99.03% F1-score, and 99.78% ROC-AUC, outperforming conventional machine learning and deep learning-based baseline models. Furthermore, FCA-based heatmap analysis provides clear temporal–spatial activation patterns of representation between healthy and Parkinsonian speech, enabling further interpretation. In conclusion, the results indicate that combining Mamba-based temporal modeling with uncertainty-aware FCA in a statistical learning architecture provides an accurate and interpretable system for automatic Parkinson’s disease detection from speech signals.