Scholarlink

<p>Early and accurate identification of Parkinson’s Disease (PD) risk seems to be an unmet challenge, particularly for diagnosing persons with sub clinical symptoms or those in early stages. Most of the existing frameworks for PD diagnosis rely on binary classification or predicting the severity of the disease by using Unified Parkinson’s Disease Rating Scale (UPDRS) score, omitting nuanced risk gradation and failing to discover PD-like healthy subgroups essential for preventive medicine. The present work proposes a multi-stage framework, combining autoencoder-based latent feature extraction, 2-tier GMM-based hierarchical clustering and adaptive model fusion to differentiate between PD, healthy, and PD-like healthy groups using voice biomarkers based on Parkinson’s Disease Similarity Index (PDSI). The proposed framework is validated using 5-fold stratified cross validation on a clinically annotated dataset of 756 voice samples, significantly skewed towards PD-positive cases. Paired t-test and Wilcoxon signed-rank statistical analysis are carried out against standard Random Forest and SVM baselines to investigate their selection correctly and reliably. The proposed model achieves a cross-validated ROC-AUC of 0.957, PR-AUC of 0.986 and F1-score of 0.945 with statistically significant improvements over ensemble and margin-based classifiers (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p&lt;0.000001\)</EquationSource> </InlineEquation>). The hierarchical approach uniquely enables detection of “PD-like healthy” individuals, thus supporting graded risk scores that can drive personalized early intervention strategies. The findings highlight PDSI as a valuable, scalable and clinically pertinent instrument for early diagnosis of PD and risk assessment and hence it creates new avenues for population wide preventative screening and incorporation into telemedicine systems.</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 Multi-stage Deep Learning and Clustering Framework for Early Parkinson’s Disease Risk Identification from Voice Biomarkers

  • 期刊论文

  • 发表日期：

    2026年5月3日

查看全文
  • Bikram Kar,
  • Bikash Kanti Sarkar,
  • Shib Sankar Sana

摘要

Early and accurate identification of Parkinson’s Disease (PD) risk seems to be an unmet challenge, particularly for diagnosing persons with sub clinical symptoms or those in early stages. Most of the existing frameworks for PD diagnosis rely on binary classification or predicting the severity of the disease by using Unified Parkinson’s Disease Rating Scale (UPDRS) score, omitting nuanced risk gradation and failing to discover PD-like healthy subgroups essential for preventive medicine. The present work proposes a multi-stage framework, combining autoencoder-based latent feature extraction, 2-tier GMM-based hierarchical clustering and adaptive model fusion to differentiate between PD, healthy, and PD-like healthy groups using voice biomarkers based on Parkinson’s Disease Similarity Index (PDSI). The proposed framework is validated using 5-fold stratified cross validation on a clinically annotated dataset of 756 voice samples, significantly skewed towards PD-positive cases. Paired t-test and Wilcoxon signed-rank statistical analysis are carried out against standard Random Forest and SVM baselines to investigate their selection correctly and reliably. The proposed model achieves a cross-validated ROC-AUC of 0.957, PR-AUC of 0.986 and F1-score of 0.945 with statistically significant improvements over ensemble and margin-based classifiers ( \(p<0.000001\) ). The hierarchical approach uniquely enables detection of “PD-like healthy” individuals, thus supporting graded risk scores that can drive personalized early intervention strategies. The findings highlight PDSI as a valuable, scalable and clinically pertinent instrument for early diagnosis of PD and risk assessment and hence it creates new avenues for population wide preventative screening and incorporation into telemedicine systems.