Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection

Researchers propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives for speech-based cognitive impairment detection. The system integrates SHAP-based token attribution, linguistic features, and an LLM reasoning pipeline to map model outputs to specific cognitive-linguistic dimensions.

• Built on the SpeechCARE-Adaptive Gating Network multimodal screening model with an F1 score of 72.11% on the NIA PREPARE benchmark.
• Utilizes a four-stage LLM reasoning pipeline powered by LLaMA-3.1-70B-Instruct to generate clinical narratives.
• Maps predictions to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence.
• Physician evaluation on 70 stratified English samples showed strong alignment with patient-level cognitive profiles.
• Achieved a System Usability Scale score of 82/100, indicating high potential for clinical workflow integration.