Towards Causal Interpretability in Deep Learning for Parkinson’s Detection from Voice Data

This research introduces a comprehensive framework for Parkinson’s Disease (PD) detection using voice recording data. We implemented and evaluated multiple deep learning models, including a baseline Convolutional Neural Network (CNN), an uncertainty-aware Monte Carlo-Dropout CNN (MCD-CNN), as well as a few-shot learning approach to address dataset size limitations. Our models achieved an accuracy over 90% in classifying PD patients using vocal biomarkers, with the ensemble model demonstrating the highest performance. We employed data augmentation techniques to address class imbalance and enhance generalization. Causal feature analysis revealed that the Noise-to-Harmonics Ratio (NHR), Recurrence Period Density Entropy (RPDE), and MDVP jitter parameters were among the most significant vocal biomarkers for PD detection, in order of estimated effect magnitude. Across deep learning models, features exhibiting the strongest absolute correlation with outputs consistently showed the largest estimated effect magnitudes. The few-shot learning approach showed promising results as well, even with limited training examples. This work demonstrates the use of causal feature analysis to validate the analysis of deep learning models, potentially enabling accessible and interpretable non-invasive screening tools.