Abstract: FR-PO0035
Artificial Intelligence (AI)-Based Extraction and Temporal Classification of Symptoms and Signs from Hemodialysis Notes
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Bergling, Karin, Renal Research Institute, New York, New York, United States
- Usvyat, Len A., Renal Research Institute, New York, New York, United States
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
Background
Symptoms and clinical signs among people receiving hemodialysis are often documented in unstructured electronic health record (EHR) notes, limiting systemic assessment of their frequency and timing. We developed a large language model (LLM) pipeline to extract, categorize, and temporally classify signs and symptoms from routine clinical documentation.
Methods
We processed 455,104 outpatient, in-center hemodialysis treatment notes from 5,000 randomly selected Fresenius Medical Care patients treated in the United States during 2024. Notes from nurses, technicians, and dietitians were merged at the treatment level. Using GPT-5-mini (OpenAI), an initial LLM pass extracted candidate symptom/sign mentions with timing, and a second LLM pass classified each mention into 16 predefined clinical and 14 temporal categories. Failed LLM calls and notes without extracted symptoms/signs were excluded. A random sample of 200 notes was reviewed by a physician to assess performance.
Results
The final analytic dataset included 127,507 treatment notes with at least one extracted symptom/sign mention, representing 4,729 patients and 192,261 total symptom/signs (mean 1.5 per treatment). See Fig. 1 for category distribution. Manual review of 200 notes evaluated LLM extraction and clinical categorization. The LLM extracted 294 candidate symptoms/signs; 10 were judged incorrectly categorized or clinically non-relevant, corresponding to 96.6% correct classification (284/294). Six relevant symptoms/signs were missed, corresponding to 98.0% extraction sensitivity (294/300). Temporal classification accuracy was 98.8% (238/241).
Conclusion
An LLM-based pipeline can structure symptoms, signs, and timing from dialysis EHR notes at scale, enabling longitudinal assessment and quality improvement in hemodialysis care.
Acknowledgment
Large language models were used as part of the workflow described in this study and to assist with abstract preparation. The authors take full responsibility for the accuracy and final content of the abstract.
Distribution of symptom and sign categories in EHR notes with extracted symptoms/signs. Categories >4% are labeled on the pie chart, while categories 0.1- 4% are displayed in the legend. Categories <0.1% are not labeled individually.
Funding
- Commercial Support – Fresenius Medical Care