MERIT: Multi-agent EHR-grounded Risk Identification and Tailored Care for Early Cognitive Decline in Primary Care
- 1Brigham and Women's Hospital/Harvard Medical School
Abstract
Background: Dementia is frequently underdiagnosed in primary care, with fewer than half of cases identified and only 8-20% detected at early stages. Our objective is to develop an Agentic AI-powered clinical decision support systems (CDSS) maintaining the patient's virtual cognitive status and provide physicians with ready-to-use summaries and next step suggestions. Methods: We developed MERIT (Multi-agent EHR-grounded Risk Identification and Tailored Care for Early Cognitive Decline in Primary Care) via participatory co-design with primary care provider. It maintains a virtual cognitive status based on the patient's electronic health records (EHR) and conversation with the patient and provide clinical decision support based on it. Retrieval augmented generation (RAG) was employed by the system to retrieve dementia-related risk factors from multimodal EHR. Additionally, we used reflection Agentic AI design pattern to improve quality of AI Agent generated text. We used one PCP's patient data to optimize the system, then used patients from another four PCPs for testing. We mainly used expert manual evaluation based on 5-point Likert-Scales. For readability of patient summary and personalized education material, besides manual evaluation, we also employed widely used Flesch Reading Ease and Flesch We-Kincaid Grade Level. Results: MERIT has four modulars: 1) patient conversation agent; 2) patient virtual cognitive status maintainer; 3) personalized education provider; and 4) decision support provider. Our evaluation focused on the provider-facing part, which relies on EHR without involving any patient conversation. The evaluation on the testing data demonstrates good performance. On a 5-point Likert scale, in the test set, patient virtual cognitive status summaries received average scores of 4.0 (SD 0.74) for correctness, 4.2 (SD 0.59) for completeness, 4.3 (SD 0.59) for readability, and 4.1 (SD 0.71) for clinical usefulness; clinical decision support received 3.9 (SD 0.87) for correctness, 4.3 (SD 0.66) for completeness, and 4.3 (SD 0.64) for adoptability; and the personalized education material received average ratings of 4.3 (SD 0.62) for completeness, 4.5 (SD 0.60) for readability, 4.3 (SD 0.78) for usefulness, and 4.4 (SD 0.49) for empathy. Automated readability assessment showed that patient summaries had a mean Flesch Reading Ease score of 18.3 (SD 4.9) and a Flesch-Kincaid Grade Level of 13.7 (SD 0.7), consistent with college-level text, whereas personalized patient education had a Flesch Reading Ease score of 50.1 (SD 7.1) and a Flesch-Kincaid Grade Level of 11.7 (SD 1.5), which is high-school student understandable text. Conclusion: Our developed system MERIT can maintain a virtual cognitive status based on the patient's multimodal EHR, provide the patient with clear and readable personalized education materials, and provide the PCP with ready-to-use cognitive summaries and decision support with high adoptability. By integrating our CDSS into Epic in the near future, our developed system will largely benefit older adults with cognitive decline and slow down the progression.
Related
- Mild Cognitive Impairment Identification in Geriatric and Primary Care Clinics and Driving Assessment Outcomes.
- CSF APOE Secondary Glycosylation Predicts Cognitive Decline in Mild Cognitive Impairment
- Interventions to reduce cognitive decline post-stroke: an umbrella review
- Searching for the Ideal Selection of Cognitive Screening Tasks for Early Detection of Cognitive Decline in Older Individuals With Low Literacy
- Childhood trauma and cognitive decline in women: an empirical analysis of cognitive vulnerability in later life
BibTeX
@misc{du2026merit,
title = {{MERIT: Multi-agent EHR-grounded Risk Identification and Tailored Care for Early Cognitive Decline in Primary Care}},
author = {X. Du and Z. Lyu and M. A. Libre and M. J. Reynolds and D. M. Rodriguez and Z. E. Liu and B. Q. Lian and L. Zhou},
howpublished = {medRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.02.26367125},
url = {https://www.medrxiv.org/content/10.64898/2026.11.02.26367125v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - MERIT: Multi-agent EHR-grounded Risk Identification and Tailored Care for Early Cognitive Decline in Primary Care
AU - X. Du
AU - Z. Lyu
AU - M. A. Libre
AU - M. J. Reynolds
AU - D. M. Rodriguez
AU - Z. E. Liu
AU - B. Q. Lian
AU - L. Zhou
PY - 2026
DA - 2026/11/02
PB - medRxiv
DO - 10.64898/2026.11.02.26367125
UR - https://www.medrxiv.org/content/10.64898/2026.11.02.26367125v1
AB - Background: Dementia is frequently underdiagnosed in primary care, with fewer than half of cases identified and only 8-20% detected at early stages. Our objective is to develop an Agentic AI-powered clinical decision support systems (CDSS) maintaining the patient's virtual cognitive status and provide physicians with ready-to-use summaries and next step suggestions. Methods: We developed MERIT (Multi-agent EHR-grounded Risk Identification and Tailored Care for Early Cognitive Decline in Primary Care) via participatory co-design with primary care provider. It maintains a virtual cognitive status based on the patient's electronic health records (EHR) and conversation with the patient and provide clinical decision support based on it. Retrieval augmented generation (RAG) was employed by the system to retrieve dementia-related risk factors from multimodal EHR. Additionally, we used reflection Agentic AI design pattern to improve quality of AI Agent generated text. We used one PCP's patient data to optimize the system, then used patients from another four PCPs for testing. We mainly used expert manual evaluation based on 5-point Likert-Scales. For readability of patient summary and personalized education material, besides manual evaluation, we also employed widely used Flesch Reading Ease and Flesch We-Kincaid Grade Level. Results: MERIT has four modulars: 1) patient conversation agent; 2) patient virtual cognitive status maintainer; 3) personalized education provider; and 4) decision support provider. Our evaluation focused on the provider-facing part, which relies on EHR without involving any patient conversation. The evaluation on the testing data demonstrates good performance. On a 5-point Likert scale, in the test set, patient virtual cognitive status summaries received average scores of 4.0 (SD 0.74) for correctness, 4.2 (SD 0.59) for completeness, 4.3 (SD 0.59) for readability, and 4.1 (SD 0.71) for clinical usefulness; clinical decision support received 3.9 (SD 0.87) for correctness, 4.3 (SD 0.66) for completeness, and 4.3 (SD 0.64) for adoptability; and the personalized education material received average ratings of 4.3 (SD 0.62) for completeness, 4.5 (SD 0.60) for readability, 4.3 (SD 0.78) for usefulness, and 4.4 (SD 0.49) for empathy. Automated readability assessment showed that patient summaries had a mean Flesch Reading Ease score of 18.3 (SD 4.9) and a Flesch-Kincaid Grade Level of 13.7 (SD 0.7), consistent with college-level text, whereas personalized patient education had a Flesch Reading Ease score of 50.1 (SD 7.1) and a Flesch-Kincaid Grade Level of 11.7 (SD 1.5), which is high-school student understandable text. Conclusion: Our developed system MERIT can maintain a virtual cognitive status based on the patient's multimodal EHR, provide the patient with clear and readable personalized education materials, and provide the PCP with ready-to-use cognitive summaries and decision support with high adoptability. By integrating our CDSS into Epic in the near future, our developed system will largely benefit older adults with cognitive decline and slow down the progression.
ER -