Scholarlink

Purpose <p>Parkinson’s disease (PD) is a complex neurodegenerative disorder influenced by genetic, clinical, and lifestyle factors. While machine learning (ML) has demonstrated potential in supporting PD diagnosis, existing approaches often rely on single-modality data and lack interpretability for individualized risk estimation. This study aimed to develop and validate an interpretable multimodal ML framework that enhances PD diagnosis and genetic subtype classification, supporting precision health and clinical decision-making.</p> Methods <p>We integrated clinical assessments, voice-derived acoustic features, and genetic variant data to construct a multimodal prediction framework. Random Forest, XGBoost, and LightGBM models were trained on clinical and voice data, while a separate classifier was trained on genetic variants for subtype identification. Stacked generalization was employed using logistic regression and support vector machine (SVM) as meta-learners. Final predictions were aggregated using weighted ensemble stacking, with AUC-informed contributions. Hierarchical clustering was applied to genetic data to explore subtype patterns. Model interpretability was examined using SHAP analysis.</p> Results <p>The proposed ensemble model achieved strong predictive performance, with AUCs exceeding 0.96 for PD diagnosis. SHAP analysis highlighted clinically relevant predictors including UPDRS scores, tremor severity, and voice frequency features. Genetic clustering revealed subtype-specific variants, including GBA1 and SNCA (linked to cognitive decline) and LRRK2 (associated with early-onset PD).</p> Conclusions <p>This interpretable and integrative framework provides accurate PD prediction and meaningful genetic subtype classification. It offers a practical foundation for AI-driven clinical decision support, aligning with personalized care initiatives and advancing the role of health technology in neurodegenerative disease management.</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

Multimodal machine learning for Parkinson’s disease diagnosis and genetic subtyping

  • 期刊论文

  • 发表日期：

    2025年10月1日

查看全文
  • Jingcheng Lu,
  • Doan Khanh Tran,
  • Xuan Lam Bui,
  • Nguyen Quoc Khanh Le,
  • Matthew Chin Heng Chua

摘要

Purpose

Parkinson’s disease (PD) is a complex neurodegenerative disorder influenced by genetic, clinical, and lifestyle factors. While machine learning (ML) has demonstrated potential in supporting PD diagnosis, existing approaches often rely on single-modality data and lack interpretability for individualized risk estimation. This study aimed to develop and validate an interpretable multimodal ML framework that enhances PD diagnosis and genetic subtype classification, supporting precision health and clinical decision-making.

Methods

We integrated clinical assessments, voice-derived acoustic features, and genetic variant data to construct a multimodal prediction framework. Random Forest, XGBoost, and LightGBM models were trained on clinical and voice data, while a separate classifier was trained on genetic variants for subtype identification. Stacked generalization was employed using logistic regression and support vector machine (SVM) as meta-learners. Final predictions were aggregated using weighted ensemble stacking, with AUC-informed contributions. Hierarchical clustering was applied to genetic data to explore subtype patterns. Model interpretability was examined using SHAP analysis.

Results

The proposed ensemble model achieved strong predictive performance, with AUCs exceeding 0.96 for PD diagnosis. SHAP analysis highlighted clinically relevant predictors including UPDRS scores, tremor severity, and voice frequency features. Genetic clustering revealed subtype-specific variants, including GBA1 and SNCA (linked to cognitive decline) and LRRK2 (associated with early-onset PD).

Conclusions

This interpretable and integrative framework provides accurate PD prediction and meaningful genetic subtype classification. It offers a practical foundation for AI-driven clinical decision support, aligning with personalized care initiatives and advancing the role of health technology in neurodegenerative disease management.