Scholarlink

This study presents a machine learning pipeline for speech-based Alzheimer’s disease detection using the ADReSSo dataset. We developed a 122-feature extraction framework spanning three domains: spectro-acoustic features capturing voice quality through energy-based vocal effort measures, mel-frequency cepstral modeling, and harmonic-spectral analysis; temporal and voice activity detection features quantifying pause patterns, speech-silence dynamics, and energy distribution statistics; and conversational interaction features analyzing turn-taking patterns, interference dynamics, and communication flow disruption. An ensemble feature selection strategy integrating six algorithms reduced dimensionality to 30 clinically interpretable features while preserving discriminative power. Among five classifiers tested, logistic regression achieved optimal performance with 82.4% accuracy and 84.2% F1-score on the validation set, with independent blind test confirming robust generalization (81% accuracy, 81% F1-score). Conversational measures, particularly turn balance, emerged as the strongest discriminators, indicating that conversational dynamics provide more reliable biomarkers than traditional acoustic features for early AD detection.

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

Thirty Features, Six Votes: Ensemble Selection of Acoustic and Conversational-Interaction Features for Alzheimer’s Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Othmane Elazri,
  • Hanae Belmajdoub,
  • Khalid Minaoui

摘要

This study presents a machine learning pipeline for speech-based Alzheimer’s disease detection using the ADReSSo dataset. We developed a 122-feature extraction framework spanning three domains: spectro-acoustic features capturing voice quality through energy-based vocal effort measures, mel-frequency cepstral modeling, and harmonic-spectral analysis; temporal and voice activity detection features quantifying pause patterns, speech-silence dynamics, and energy distribution statistics; and conversational interaction features analyzing turn-taking patterns, interference dynamics, and communication flow disruption. An ensemble feature selection strategy integrating six algorithms reduced dimensionality to 30 clinically interpretable features while preserving discriminative power. Among five classifiers tested, logistic regression achieved optimal performance with 82.4% accuracy and 84.2% F1-score on the validation set, with independent blind test confirming robust generalization (81% accuracy, 81% F1-score). Conversational measures, particularly turn balance, emerged as the strongest discriminators, indicating that conversational dynamics provide more reliable biomarkers than traditional acoustic features for early AD detection.