Scholarlink

Purpose <p>Current deep learning approaches for Parkinson's Disease (PD) diagnosis lack causal interpretability, limiting clinical trust and biomarker discovery. This study establishes a causal framework to answer two critical questions: 1) Which acoustic features causally drive PD diagnosis? 2) Which temporal phases of speech production are most diagnostically significant?</p> Methods <p>We developed a counterfactual causal inference framework based on the Rubin Causal Model (RCM), implemented through a hybrid TCN-Transformer predictive model (AUC = 0.952). Using Average Causal Effect (ACE) analysis, we quantified the causal impact of 10 global acoustic features. Temporal causality was assessed through segmented masking interventions across speech signals.</p> Results <p>Fundamental frequency (F0) features emerged as the strongest causal drivers, with F0 mean showing the highest ACE (0.087, <i>p</i> &lt; 0.001). Temporal analysis revealed 73.2% of diagnostic information resides in the initial 20% of phonation (Voice Onset Difficulty Index, VODI: PD = 64.4% vs. HC = 23.1%, <i>p</i> &lt; 0.001). Our model significantly outperformed seven baselines, achieving 0.952 AUC (Δ + 2.9% vs. best competitor).</p> Conclusions <p>This is the first study to apply a counterfactual causal inference framework to quantitatively establish F0 characteristics as causal PD biomarkers and identify phonation onset as the critical diagnostic window. The Voice Onset Difficulty Index (VODI) provides a validated metric for assessing speech initiation impairment. Our framework advances PD diagnosis from correlational prediction to causal explanation, offering clinically actionable insights.Importantly, these causal estimates are model-based counterfactual effects inferred from observational data; prospective interventional studies are needed to confirm clinical causality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A counterfactual inference framework for Parkinson’s disease diagnosis: uncovering the causal role of phonation onset in speech

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Leiyong Guo,
  • Yu Li

摘要

Purpose

Current deep learning approaches for Parkinson's Disease (PD) diagnosis lack causal interpretability, limiting clinical trust and biomarker discovery. This study establishes a causal framework to answer two critical questions: 1) Which acoustic features causally drive PD diagnosis? 2) Which temporal phases of speech production are most diagnostically significant?

Methods

We developed a counterfactual causal inference framework based on the Rubin Causal Model (RCM), implemented through a hybrid TCN-Transformer predictive model (AUC = 0.952). Using Average Causal Effect (ACE) analysis, we quantified the causal impact of 10 global acoustic features. Temporal causality was assessed through segmented masking interventions across speech signals.

Results

Fundamental frequency (F0) features emerged as the strongest causal drivers, with F0 mean showing the highest ACE (0.087, p < 0.001). Temporal analysis revealed 73.2% of diagnostic information resides in the initial 20% of phonation (Voice Onset Difficulty Index, VODI: PD = 64.4% vs. HC = 23.1%, p < 0.001). Our model significantly outperformed seven baselines, achieving 0.952 AUC (Δ + 2.9% vs. best competitor).

Conclusions

This is the first study to apply a counterfactual causal inference framework to quantitatively establish F0 characteristics as causal PD biomarkers and identify phonation onset as the critical diagnostic window. The Voice Onset Difficulty Index (VODI) provides a validated metric for assessing speech initiation impairment. Our framework advances PD diagnosis from correlational prediction to causal explanation, offering clinically actionable insights.Importantly, these causal estimates are model-based counterfactual effects inferred from observational data; prospective interventional studies are needed to confirm clinical causality.