SCH: Fair and Interpretable Prediction of Alzheimer's Disease and MCI using LLMs

This proposal aims to develop high-performing, fair, and interpretable large language models (LLMs) to detect Alzheimer's disease (AD) and its precursor stages. In doing so, its overarching goal is to advance healthcare through the innovative and responsible use of LLMs to recognize AD, mild cognitive impairment (MCI), and preclinical AD from transcribed unstructured speech samples, prioritizing equity and fairness in their deployment. Evidence has demonstrated that language is impaired early in AD, including in the MCI and even prodromal AD stage. Moreover, recent natural language processing (NLP) approaches have shown that informative markers of this impairment can be captured from spontaneous speech samples. LLMs, a new cornerstone of NLP, have shown promise at addressing healthcare tasks, but concerns exist regarding their potential biases and poor explainability. We will test the use of LLMs to identify early stages of the AD trajectory from transcribed unstructured speech samples, making equity, fairness, and explainability a central piece of our aims. We hypothesize that the use of fair, explainable LLMs for early AD detection may overcome data constraints and interpretability issues common while enabling effective low-resource learning for this purpose. We will test this hypothesis through three aims: (1) building on our shared expertise and prior work, we will design LLM-based models for detecting MCI and dementia and explore their use in detecting the earliest preclinical stage; (2) we will enhance model fairness and interpretability through innovative demonstration example selection and domain-aligned instruction tuning techniques, and develop new methods for explaining LLM output in healthcare settings; and (3) we will validate our methods by assessing their alignment with clinical parameters related to AD and MCI. We will also perform a pioneering exploration of the extent to which our findings are repeatable beyond English settings. Expected outcomes include groundbreaking generalizable insights in NLP and AI fairness and interpretability, and domain-specific advances that could help ensure equitable access to AD and MCI detection. This may lay groundwork for a low-cost, scalable, non-invasive alternative for cognitive health screening and reduce healthcare disparities via more equitable access to high-quality cognitive health assessment. At an economic scale, this work may reduce long-term burden on the healthcare system and improve quality of life, benefitting individuals and society. RELEVANCE (See instructions): It is critical to mark the earliest changes in brain function among those on the Alzheimer's disease (AD) trajectory, and subtle language changes may be a sensitive, non-invasive screening tool. We will harness fair, responsible, interpretable AI to analyze natural speech patterns as a marker of early AD. This will advance understanding of language changes in early AD and lay groundwork for equitable access to AD assessment, potentially reducing healthcare disparities and alleviating burden on the healthcare system.