Abstract: SA-PO0054
A Large Language Model (LLM)-Based Approach to Extract Kidney Pathology Report Data from Electronic Health Records
Session Information
- Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
October 24, 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
- Tajerian, Amin, Baylor Scott & White Research Institute, Dallas, Texas, United States
- Arora, Akshay, Baylor Scott & White Research Institute, Dallas, Texas, United States
- Cheng, Mingyu, Baylor Scott & White Medical Center Temple, Temple, Texas, United States
- Ettlinger, Jason, Baylor Scott & White Research Institute, Dallas, Texas, United States
- Priest, Elisa, Baylor Scott & White Research Institute, Dallas, Texas, United States
- Bobart, Shane A., Mayo Clinic in Florida, Jacksonville, Florida, United States
- Shafi, Tariq, Baylor Scott & White Medical Center Temple, Temple, Texas, United States
Background
Kidney biopsy reports are critical for population-level glomerular disease research but remain unstructured and non-indexed in the electronic health records, requiring labor-intensive manual abstraction. We aimed to evaluate an LLM–based approach to extract key pathology data from free-text kidney biopsy reports and convert them into a structured dataset.
Methods
We obtained archived kidney biopsy reports from Baylor Scott & White from 2010 to 2025. Reports were de-identified using the deid_bert_i2b2 LLM and classified into one or more predefined diagnostic categories using a general-purpose LLM (GPT 4.1, OpenAI). Task-specific prompting and structured tool calls were used to ensure standardized, machine-readable outputs, and the model was allowed to assign multiple diagnoses to a single report when applicable.
Results
A total of 2,658 kidney biopsy reports were included. Mean patient age was 50.1 years; 52.1% were female, 22.5% identified as Black or African American, and 23.9% were Hispanic. Mean eGFR before biopsy was 46. The LLM generated a structured dataset with a median of 5 diagnoses (IQR: 3, 6) per report. The most common primary diagnoses were diabetic kidney disease (21.1%), focal segmental glomerulosclerosis (20.5%), lupus nephritis (14.7%), and crescentic GN (14.9%). 66% of reports showed some degree of tubulointerstitial fibrosis, and 76% reported some degree of arteriosclerosis.
Conclusion
These findings demonstrate that LLM-based methods, using a constrained output format and without fine-tuning, can efficiently transform unstructured kidney biopsy reports into analyzable data. This approach has the potential to substantially reduce manual abstraction burden and enable scalable integration of pathology data into population health research and precision nephrology care.
Distribution of primary kidney biopsy diagnoses, after excluding tumor and transplanted kidneys.
| Diagnosis | N (%) | Diagnosis (continued) | N (%) | Diagnosis (continued) | N (%) |
| Diabetic kidney disease | 486 (19.7%) | Acute interstitial nephritis | 76 (3.1%) | IgM Glomerulopathy | 16 (0.6%) |
| Lupus nephritis | 355 (14.4%) | Minimal change disease | 69 (2.8%) | Fibrillary glomerulonephritis | 15 (0.6%) |
| IgA nephropathy | 224 (9.1%) | Membranoproliferative glomerulonephritis | 66 (2.7%) | C1q nephropathy | 15 (0.6%) |
| Focal segmental glomerulosclerosis | 202 (8.2%) | Immune Complex Mediated Glomerulonephritis | 66 (2.7%) | Chronic interstitial nephritis | 10 (0.4%) |
| Membranous nephropathy | 126 (5.1%) | Arterionephrosclerosis | 48 (1.9%) | C3 glomerulonephritis | 9 (0.4%) |
| Global glomerulosclerosis | 124 (5.0%) | Thrombotic microangiopathy | 44 (1.8%) | Anti-GBM nephritis | 6 (0.2) |
| Acute tubular necrosis | 109 (4.4%) | Crescentic Glomerulonephritis | 41 (1.7%) | Genetic kidney disease (Alport's, Fabry's) | 5 (0.2) |
| Pauci-immune glomerulonephritis/ANCA vasculitis | 92 (3.7%) | Amyloidosis | 34 (1.4%) | Diffuse Proliferative Glomerulonephritis | 4 (0.2) |
| Thin Basement Membrane | 84 (3.4%) | Tubulointerstitial Nephropathy and fibrosis | 33 (1.3%) | Other, Normal or Non-diagnostic | 108 (4.4%) |