Kidney Week

Abstract: TH-OR027

Large Language Models for CKD Guideline-Directed Medical Therapy

Session Information

• Artificial Intelligence in Kidney Care: From Pathology to Prediction
  October 22, 2026 | Location: Room 711, Convention Center
  Abstract Time: 05:50 PM - 06:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Badrouchi, Samarra, Brigham and Women's Hospital, Boston, United States

Samarra Badrouchi, MD, MMSc

• Lo, Kevin Bryan, Brigham and Women's Hospital, Boston, Massachusetts, United States

Kevin Bryan Lo, MD

• Helseth, Ragnhild, Brigham and Women's Hospital, Boston, Massachusetts, United States

Ragnhild Helseth, MD, PhD

• Cruz Solbes, Ana Sofia, Brigham and Women's Hospital, Boston, Massachusetts, United States

Ana Sofia Cruz Solbes, MD

• Marti, Pablo Miki, Brigham and Women's Hospital, Boston, Massachusetts, United States

Pablo Miki Marti, MSc, PhD

• Claggett, Brian, Brigham and Women's Hospital, Boston, Massachusetts, United States

Brian Claggett, MS, PhD

• Vaduganathan, Muthiah, Brigham and Women's Hospital, Boston, Massachusetts, United States

Muthiah Vaduganathan, MD, MPH

• Solomon, Scott D., Brigham and Women's Hospital, Boston, Massachusetts, United States

Scott D. Solomon, MD

• McGrath, Martina M., Brigham and Women's Hospital, Boston, Massachusetts, United States

Martina M. McGrath, MBChB, FASN

• Mc Causland, Finnian R., Brigham and Women's Hospital, Boston, Massachusetts, United States

Finnian R. Mc Causland, MBChB, MS, FASN

• Cunningham, Jonathan W., Brigham and Women's Hospital, Boston, Massachusetts, United States

Jonathan W. Cunningham, MD, MPH

Background

Patients with chronic kidney disease (CKD) are often undertreated with guideline-directed medical therapies (GDMT) that slow kidney decline and reduce cardiovascular risk. Guideline implementation requires clinicians to synthesize patients' laboratory trends, contraindications, and prior intolerances. Large language models (LLMs) may help synthesize electronic health record (EHR) data to support GDMT optimization.

Methods

We developed and validated an automated workflow that extracts EHR information and uses an LLM to generate evidence-based patient-specific CKD medication recommendations. We analyzed patients with upcoming nephrology, endocrinology, or primary care visits from April-July 2026 at Brigham and Women’s Hospital. CKD was defined by eGFR <60 mL/min/1.73m² and/or albuminuria ≥30mg/g, sustained for ≥90 days. We excluded patients on dialysis, kidney transplant recipients, age ≥85, or eGFR <20mL/min/1.73m². Structured and narrative EHR data were extracted from the clinical data warehouse and processed through a purpose-built secure GPT-5.4 workflow. Two clinicians reviewed 200 randomly-selected messages with ≥1 recommendation for appropriateness and safety.

Results

Among 20,470 screened patients, 2,996 (15%) were eligible; 945 (31.5%) had ≥1 LLM-generated recommendation. Eligible patients had mean age of 68.7 years; 1,652 (55.1%) had type 2 diabetes, 2,698 (90.1%) had hypertension, and 528 (17.6%) had heart failure. Mean BMI was 29.9 kg/m², and 1,079/2,544 (42.4%) had obesity. The LLM recommended SGLT2i in 543 patients (57.5%), ACEi/ARB in 271 (28.7%), nsMRA in 142 (15.0%), GLP-1RA in 93 (9.8%), and sMRA in 7 (0.7%). In validation, 200/200 recommendations were appropriate and 199/200 (99.5%) were safe; the single borderline case recommended SGLT2i in a patient with chronic Foley/neurogenic bladder and prior urinary tract infection, which the LLM noted in the message's safety section.

Conclusion

An automated LLM-based workflow identified CKD GDMT opportunities in one-third of patients, with high appropriateness and safety. Sharing LLM-generated recommendations with clinicians could improve GDMT utilization.