Language-Based Dementia Classification Should Consider Model Cognition for Interpretability

Yui Ishihara, Michelle Cohn, Kartik Patwari, Alyssa Weakley, Chen-Nee Chuah

Keywords: Position, Dementia Detection, LLMs, Interpretability, Model Cognition

TL;DR: This position paper argues that language-based dementia detection should include probing model cognition through structured, fine-grained input-output analyses, which will enable clearer understanding of how ML systems reason in this task.

Abstract: Current approaches to dementia detection in machine learning based on language often treat the task as an end-to-end binary classification problem, directly classifying a person's audio or transcripts to a final diagnostic label (e.g., dementia or cognitively normal). While these prior techniques can be effective for accuracy, because predictions are made in a single step, this binary approach overlooks the progressive nature of cognitive decline and lacks the interpretive analyses that clinicians rely on in clinical settings. Furthermore, these approaches face interpretability limitations, particularly in terms of the linguistic features the models are focusing on. **This position paper argues that language-based dementia detection research should be re-framed to include cognition-based reasoning, specifically by probing model cognition through structured, fine-grained input–output analyses, which allow clearer understanding of how ML systems reason in this task.** This new direction will help advance current ML models towards dementia detection frameworks that are more interpretable and clinically trustworthy.

Submission Number: 24