BACKGROUND:
Family members and caregivers play a vital role in the dementia diagnostic pathway, often being the first to notice subtle changes and later providing essential information through informant reports. These accounts capture everyday cognitive, behavioural, and functional changes that may not be apparent during clinical assessment. Despite their value, such reports are often inconsistently collected and underutilised due to limited appointment time and variability in clinician interpretation. Advances in Large Language Models (LLMs) offer new opportunities to automate and standardise the analysis of this information. This study investigates whether caregiver-provided information (known as a collateral history) can be automatically analysed using an LLM to support dementia diagnosis.
METHODS:
Forty structured interviews will be conducted with family members and caregivers of individuals with Mild Cognitive Impairment (MCI) or dementia. Interviews will capture observations and insights into early cognitive, behavioural, and functional changes. Transcripts will be anonymised and analysed by two expert raters (a consultant neurologist and an old age psychiatrist) to assess inter-rater reliability, followed by analysis using an LLM (LLaMA 3.2-3B Instruct). The model will be evaluated on its ability to: (1) identify evidence of cognitive impairment, (2) describe domain-specific deficits, (3) suggest differential diagnoses, and (4) detect signs of caregiver stress. LLM outputs will be compared with clinician analyses to assess concordance and thematic accuracy.
RESULTS:
Data collection and analysis are ongoing. Early pilot work has focused on refining the interview structure, optimising LLM prompts, and ensuring clinical relevance in model outputs. Preliminary findings suggest the LLM can extract structured, clinically meaningful information from caregiver data.
CONCLUSION:
This study will assess whether LLMs can reliably interpret caregiver-reported observations to support earlier and more consistent dementia diagnosis. Integrating AI into diagnostic workflows could enhance efficiency, reduce clinician workload, and improve access to timely, remote assessments.