Development and Validation of a Machine Learning Method Using Vocal Biomarkers for Identifying Frailty in...

Area of Science:

• Gerontology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Current frailty assessment methods, including the frailty phenotype and frailty index, have limitations for clinical application.
• Standardized methods for measuring frailty are still lacking, highlighting the need for innovative approaches.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) classification model for predicting frailty status in community-dwelling older adults using vocal biomarkers.
• To leverage voice recordings from the picture description task (PDT) for noninvasive frailty assessment.

Main Methods:

• Recruited 127 participants aged 50+ and collected clinical data using the Comprehensive Geriatric Assessment scale.
• Collected voice recordings via tablet during the Korean PDT, preprocessing audio to remove noise.
• Developed three AI models: SpeechAI (voice data only), DemoAI (demographic data only), and DemoSpeechAI (combined data).

Main Results:

• The SpeechAI model achieved 80.4% accuracy and an AUC of 0.89, outperforming demographic-only models (67.96% accuracy, AUC 0.74).
• The combined DemoSpeechAI model showed superior performance with 85.6% accuracy and an AUC of 0.93.
• SpeechAI significantly outperformed traditional acoustic feature methods (AUC 0.89 vs. 0.57-0.66).

Conclusions:

• Vocal biomarkers analyzed by deep learning-based AI models can effectively predict frailty in community-dwelling older adults.
• Speech-based AI models provide a noninvasive and scalable method for frailty detection.
• These AI models can potentially streamline frailty assessments in clinical and community settings.