Scholarlink

<p>Most voice-based Parkinson’s disease studies evaluated a narrow set of models on fixed feature spaces, which limited generalizability and obscured feature–model synergies. Many also neglected fold-wise fitting and runtime measurement, increasing the risk of data leakage and hindering telehealth deployment. Therefore, the study developed a leak-safe, runtime-aware factorial benchmark that crossed nine representation strategies with twelve classifiers, giving 108 pipelines under stratified 10-fold cross-validation on a 22-feature sustained-vowel dataset. All transformations were fit strictly within training folds, class imbalance was addressed with in-fold SMOTE, and four endpoints, accuracy, F1, AUC, and per-fold runtime, were jointly analyzed across classical machine learning, ensembles, and interaction-aware deep networks, explicitly profiling accuracy–latency trade-offs relevant to point-of-care screening. The results showed that simple, information-preserving filters paired with interaction-aware learners dominated the Pareto frontier. The top pipeline, Variance plus Deep &amp; Cross, achieved mean accuracy 94.37% ± 4.76, F1 94.26% ± 4.83, and AUC 0.9852 ± 0.0234, with 11.73 s per fold. A low-latency alternative, Correlation plus KNN, delivered 92.32% accuracy, 92.15% F1, and 0.9799 AUC in about 0.15 s per fold, offering a strong speed–accuracy trade-off. Additional high performers included PCA plus Wide &amp; Deep (92.82% accuracy, 92.64% F1, 0.9560 AUC) and Correlation plus Deep &amp; Cross (92.34% accuracy, 92.20% F1, 0.9797 AUC). Manual top-k raw features and coarse statistical aggregates underperformed across the 108 pipelines. In conclusion, optimizing representation–learner synergy is pivotal for robust, scalable PD voice screening, with variance or correlation filters plus interaction-aware deep models yielding state-of-the-art discrimination and Correlation plus KNN offering an ultra-fast option for resource-constrained settings.</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

Optimizing Voice-Based Feature Engineering and Machine Learning Pipelines for Early Parkinson’s Disease Detection

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Fatin A. Elhaj,
  • Taqwa A. Alhaj,
  • Jahanzeb Sheikh,
  • Sania Tanvir,
  • Kashif Shaikh,
  • Tan Tian Swee,
  • Madeeha Sadia,
  • Michael Loong Peng Tan,
  • Ammar Ahmed Mohamed Elhaj,
  • Errachid El Salhi Abdelhamid

摘要

Most voice-based Parkinson’s disease studies evaluated a narrow set of models on fixed feature spaces, which limited generalizability and obscured feature–model synergies. Many also neglected fold-wise fitting and runtime measurement, increasing the risk of data leakage and hindering telehealth deployment. Therefore, the study developed a leak-safe, runtime-aware factorial benchmark that crossed nine representation strategies with twelve classifiers, giving 108 pipelines under stratified 10-fold cross-validation on a 22-feature sustained-vowel dataset. All transformations were fit strictly within training folds, class imbalance was addressed with in-fold SMOTE, and four endpoints, accuracy, F1, AUC, and per-fold runtime, were jointly analyzed across classical machine learning, ensembles, and interaction-aware deep networks, explicitly profiling accuracy–latency trade-offs relevant to point-of-care screening. The results showed that simple, information-preserving filters paired with interaction-aware learners dominated the Pareto frontier. The top pipeline, Variance plus Deep & Cross, achieved mean accuracy 94.37% ± 4.76, F1 94.26% ± 4.83, and AUC 0.9852 ± 0.0234, with 11.73 s per fold. A low-latency alternative, Correlation plus KNN, delivered 92.32% accuracy, 92.15% F1, and 0.9799 AUC in about 0.15 s per fold, offering a strong speed–accuracy trade-off. Additional high performers included PCA plus Wide & Deep (92.82% accuracy, 92.64% F1, 0.9560 AUC) and Correlation plus Deep & Cross (92.34% accuracy, 92.20% F1, 0.9797 AUC). Manual top-k raw features and coarse statistical aggregates underperformed across the 108 pipelines. In conclusion, optimizing representation–learner synergy is pivotal for robust, scalable PD voice screening, with variance or correlation filters plus interaction-aware deep models yielding state-of-the-art discrimination and Correlation plus KNN offering an ultra-fast option for resource-constrained settings.