From Black-Box to Clinical Insight: A Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection

2026-06-26

AI Summary

This study addresses the lack of clinical interpretability in Transformer-based models for detecting speech-related cognitive impairments by proposing a multi-stage interpretable framework. The approach uniquely integrates SHAP token attribution, cognitive-linguistic theory–driven feature engineering, and a four-step reasoning chain powered by LLaMA-3.1-70B-Instruct, combined with the SpeechCARE-Adaptive Gating Network multimodal model. This integration transforms opaque model predictions into clinically comprehensible cognitive-linguistic narratives. Evaluated on the NIA PREPARE benchmark, the method achieves an F1 score of 72.11%. Clinician assessments confirm high alignment between the system’s outputs and patients’ cognitive profiles, and the system receives a usability score of 82 out of 100, demonstrating a transparent mapping from model decisions to clinically relevant cognitive dimensions.

Abstract

Speech-based cognitive impairment detection offers a noninvasive, accessible alternative to costly biomarker assays, yet transformer-based models remain clinically uninterpretable. We propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives by integrating SHapley Additive exPlanations (SHAP)-based token attribution, theory-informed linguistic features, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. Built on the SpeechCARE-Adaptive Gating Network multimodal screening model (F1 = 72.11% on the NIA PREPARE benchmark), the framework maps model outputs to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence. Physician evaluation on 70 stratified English samples demonstrated strong alignment with patient-level cognitive profiles, and a System Usability Scale score of 82/100 indicated high potential for clinical workflow integration.

Problem

• cognitive impairment detection
• speech-based assessment
• model interpretability
• clinical explainability
• transformer models

Innovation

• explainable AI
• speech-based cognitive impairment detection
• SHAP attribution
• large language model reasoning
• cognitive-linguistic dimensions