Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing
• Health Informatics

Background:

• Artificial intelligence (AI) and deep learning (DL) show promise for audio-based voice and speech biomarkers in healthcare.
• The 'black-box' nature of DL models raises ethical and regulatory concerns, limiting clinical adoption.
• Explainable AI (XAI) offers a solution by providing understandable explanations for AI decisions, enhancing trustworthiness.

Purpose of the Study:

• To systematically review XAI methods used to explain DL models in clinical audio applications.
• To identify XAI-informed insights and assess their clinical applicability and stakeholder relevance.
• To propose recommendations for future clinical audio XAI design.

Main Methods:

• Systematic literature review following PRISMA guidelines.
• Searched six databases (IEEE Xplore, ACM, Scopus, PubMed, Web of Science, Nature) from January 2015 to February 2025.
• Thematic synthesis of 30 eligible studies based on explainability methods, clinical domains, and validation strategies.

Main Results:

• A variety of XAI methods were employed, including gradient-based, perturbation-based, and attention-based techniques.
• Applications covered diverse clinical areas such as voice disorders, neurodegenerative diseases, and psychiatric conditions.
• Most studies used qualitative interpretation with limited quantitative validation and lacked stakeholder evaluations.

Conclusions:

• Current XAI in clinical voice analysis suffers from inadequate validation and domain-specific design.
• There's a need for validated, audio-aware, and stakeholder-centered XAI approaches for clinical trust.
• Future XAI development must address limitations in validation and ensure alignment with clinical needs.