Scholarlink

<p>The effect of Levodopa on speech in Parkinson’s disease (PD) is inconsistently reported, likely due to phenotypic heterogeneity. This study introduces a comprehensive analysis based on both statistical testing and an explainable machine learning (ML) pipeline to objectively detect and characterize speech changes in PD motor subtypes—postural instability/gait difficulty (PIGD) and tremor-dominant (TD)—in response to an acute Levodopa challenge. A machine-learning (ML) pipeline differentiated healthy controls (HCs) from PD patients (stratified into PIGD/TD subtypes) and OFF/ON medication states using speech data. Acoustic features representing respiration, phonation, articulation, prosody, and their fusion were extracted from sustained vowels and diadochokinetic (DDK) tasks in Tunisian Arabic. Support Vector Machine (SVM) and Random Forest (RF) models were optimized using nested cross-validation, with feature selection via RFECV and class rebalancing with SMOTE. Shapley Additive Explanations (SHAP) values identified key biomarkers, confirmed statistically across groups. The SVM model performed better in MCC for most binary classifiers. Subtype-driven analysis revealed that speech disorders were significantly more evident in the PIGD group than in the TD group, with articulatory abnormalities emerging as key features. For medication state, Levodopa improved speech especially PIGD articulation and TD phonation. This research introduces a dual ML pipeline monitors Levodopa’s speech effects in PD subtypes, revealing distinct medication responses. The approach is scalable and interpretable, offering a robust foundation for clinical trial stratification and personalized treatment decision support.</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

An explainable machine learning model for detecting behavioral medication effects in motor subtypes of early Parkinson’s disease based on acoustics speech signals

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Zeineb BenMessaoud,
  • Sonia BenHassen,
  • Mohamed Neji,
  • Amir Hussain,
  • Nouha Farhat,
  • Emna Smaoui,
  • Mariem Dammek,
  • Mondher Frikha,
  • Adel M Alimi,
  • Chokri Mhiri

摘要

The effect of Levodopa on speech in Parkinson’s disease (PD) is inconsistently reported, likely due to phenotypic heterogeneity. This study introduces a comprehensive analysis based on both statistical testing and an explainable machine learning (ML) pipeline to objectively detect and characterize speech changes in PD motor subtypes—postural instability/gait difficulty (PIGD) and tremor-dominant (TD)—in response to an acute Levodopa challenge. A machine-learning (ML) pipeline differentiated healthy controls (HCs) from PD patients (stratified into PIGD/TD subtypes) and OFF/ON medication states using speech data. Acoustic features representing respiration, phonation, articulation, prosody, and their fusion were extracted from sustained vowels and diadochokinetic (DDK) tasks in Tunisian Arabic. Support Vector Machine (SVM) and Random Forest (RF) models were optimized using nested cross-validation, with feature selection via RFECV and class rebalancing with SMOTE. Shapley Additive Explanations (SHAP) values identified key biomarkers, confirmed statistically across groups. The SVM model performed better in MCC for most binary classifiers. Subtype-driven analysis revealed that speech disorders were significantly more evident in the PIGD group than in the TD group, with articulatory abnormalities emerging as key features. For medication state, Levodopa improved speech especially PIGD articulation and TD phonation. This research introduces a dual ML pipeline monitors Levodopa’s speech effects in PD subtypes, revealing distinct medication responses. The approach is scalable and interpretable, offering a robust foundation for clinical trial stratification and personalized treatment decision support.