Characterizing Dementia Phenotypes from Unstructured EHR Notes with Generative AI and Interpretable Machine Learning.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Dementia, including Alzheimer's disease (AD) and behavioral-variant frontotemporal dementia (bvFTD), presents diverse symptoms challenging accurate diagnosis.
• Unstructured clinical notes in electronic health records (EHRs) contain vital diagnostic information.
• Leveraging advanced AI, like large language models (LLMs), can unlock insights from this unstructured data.

Purpose of the Study:

• To develop and validate a pipeline using LLMs for symptom phenotyping from EHR notes.
• To extract and structure clinical findings for improved dementia syndrome characterization.
• To differentiate between AD and bvFTD using extracted symptom data and machine learning.

Main Methods:

• Utilized LLMs to process over 350,000 specialty notes from UCSF.
• Developed a pipeline for extracting and structuring symptom phrases into 51 distinct symptom groups.
• Applied traditional machine learning (logistic regression) to classify patients into AD and bvFTD syndromes.

Main Results:

• Successfully identified cohorts of 122 bvFTD and 170 AD patients.
• Extracted 12,637 distinct symptom phrases, clustered into 51 groups.
• Achieved an AUC of 0.83 in differentiating AD and bvFTD, identifying key distinguishing symptoms (e.g., disinhibition for bvFTD, visuospatial issues for AD).

Conclusions:

• LLM-based information extraction combined with machine learning offers a powerful approach for dementia symptom characterization.
• This methodology shows promise for enhancing diagnostic accuracy and developing predictive models for dementia.
• Potential applications include optimizing treatment strategies and improving patient care in dementia syndromes.