1468: A Hipaa-compliant LLM for Personalized Therapeutic Decision-making in Lupus Nephritis

Background/Purpose: Lupus nephritis (LN) is the most significant driver of morbidity and mortality in patients with systemic lupus erythematosus (SLE). The ACR 2024 lupus nephritis treatment guidelines (PMID: 40127995) emphasize the importance of intensive induction of LN flares with initiation of dual and/or triple therapy. The guidelines also recommend prompt pulse steroid therapy for both proliferative and pure membranous nephritis. However, these recommendations are conditional, and the necessity for the same treatment intensity for all LN, without regard for other features on biopsy (e.g. crescents, high activity index (AI)), is less certain. Recent data indicate that an AI of <3 predicts a more favorable prognosis, supporting the concept that a subset of patients can achieve sustained remission without pulse steroids (PMID: 39521057,40180257). Therefore, an evidence-based individualized treatment approach is essential for LN patients, prompting us to create a clinical decision-making support tool to stratify patients for treating providers.

Methods: We used a local HIPAA-compliant large language model (LLM), NYUTron, trained on 10 years of data from 387,144 patients (PMID: 37286606), currently used for clinical predictive modeling in our hospital system. This model organized renal biopsy reports into tabular data of AI, chronicity indices, crescents, sclerosis, thrombi, immunofluorescence, deposits, exostin2 staining, interstitial fibrosis, tubular atrophy, and arterial intimal fibrosis. Prompt engineering involved summarizing ISN/RPS criteria and ideal table examples (Fig 1A). We reviewed renal biopsies from 1/2024 to 3/2025 with ICD-10 SLE diagnoses, comparing results with LLM output to validate accuracy.

Results: 70 renal biopsies were included, with narrative reports generated by 10 renal pathologists (Arkana labs). Of these, four had limited specimens (< 10 glomeruli) or alternate pathology (e.g., focal segmental glomerular sclerosis, acute tubular injury, or lupus podocytopathy), all correctly identified by the LLM (100% accuracy). For ISN/RPS classifications, the LLM correctly assigned A (active), C (chronic), or A/C lesions in 65 of 66 biopsies (98.5% accuracy). The LLM demonstrated 100% accuracy in reporting microthrombi (6/70 biopsies) and cellular/fibrocellular crescents (23/70 biopsies) (Table 1). Arteriolar hyalinosis was accurately identified in 69 of 70 biopsies (98.6% accuracy). This LLM's retrospective renal biopsy summarization, validated by physician review, shows the feasibility of identifying relevant features in SLE renal biopsies to aid in individualized therapy plans (Fig 1B).

Conclusion: This study highlights our LLM's accuracy in analyzing LN renal biopsies, identifying key pathological features linked to poor prognosis. Summarizing these features can facilitate personalized therapy approaches through risk stratification, ultimately improving lupus nephritis treatment outcomes with further validation. We will focus next on fine-tuning our LLM model and integrating biopsy reports with unstructured rheumatologist notes and patient outcomes, aiming to deploy this model in a future prospective, single-arm trial.