Unveiling the causal pathway of Parkinson's disease dysphonia: A voice causal generative model (VCGM) approach.

BACKGROUND: Although artificial intelligence (AI) models have demonstrated high accuracy in diagnosing Parkinson's disease (PD) from speech signals, their "black-box" nature prevents mechanistic understanding of vocal impairment, limiting clinical trust and utility. A paradigm shift from correlation-based explanation to causal reasoning is needed to unlock the potential of AI in computational medicine. METHODS: We propose the Voice Causal Generative Model (VCGM), a novel computational framework designed to infer physiologically plausible causal pathways from observational speech data. The core technical innovation of VCGM is the integration of biophysical knowledge as hard constraints within a linear non-Gaussian acyclic model. We formalize this in Theorem 1, which proves that these domain-specific hierarchical constraints guarantee the unique identifiability of the underlying causal structure, a condition unachievable by unconstrained methods. We implemented VCGM using a constrained DirectLiNGAM algorithm and conducted rigorous validation, including bootstrap analysis for stability, an ablation study, and comparison with traditional causal discovery algorithms. RESULTS: The VCGM uncovered a stable, medically plausible causal pathway for PD dysphonia. The model revealed a hierarchical cascade from disease status to physiological instability (e.g., the robust Shimmer→HNR pathway) and finally to acoustic distortion (e.g., HNR→MFCC2). VCGM reduced physiologically implausible edges by 100% compared to unconstrained LiNGAM (5 implausible edges) and GES (5 implausible edges). while a conventional SHAP-based associative model failed to provide any directional mechanistic insights. CONCLUSION: The VCGM provides a validated, "white-box" framework for deconstructing pathophysiology from complex biosignals. Its primary technical contribution is a provably identifiable modeling approach that makes causal discovery feasible and reliable in hierarchically structured domains. It marks a critical step from the associative "what" to the causal "why" in computational biomedicine, offering a blueprint for more transparent, trustworthy, and clinically insightful AI systems. To facilitate clinical translation and reproducibility, all code and data are publicly available under an open-source license.