KT-LLM: an evidence-grounded and sequence text framework for auditable kidney transplant modeling.

Area of Science:

• Nephrology and Biomedical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Real-world kidney transplantation (KT) faces challenges with inconsistent follow-up and policy adherence.
• A disconnect exists between structured data and clinical rules, impacting reporting and reproducibility.
• Existing systems lack auditable reasoning grounded in authoritative clinical guidelines.

Purpose of the Study:

• To introduce KT-LLM, an auditable orchestration layer for kidney transplantation workflows.
• To bridge sequence modeling with policy-aware reasoning for improved clinical decision-making.
• To ensure clinical decision-making is grounded in versioned, authoritative sources like Banff pathology and OPTN/SRTR policies.

Main Methods:

• Developed KT-LLM, a retrieval-augmented generation system using versioned policy documents.
• Integrated three auditable agents: SRTR-MambaSurv for survival prediction, OPTN-BlackClust for subtype identification, and Policy-Ops for compliance checks.
• Utilized de-identified OPTN and UNOS cohorts for validation.

Main Results:

• KT-LLM demonstrated superior evidence attribution and predictive calibration compared to baselines.
• The system successfully identified distinct subgroups among Black recipients, aligning with outcome heterogeneity.
• KT-LLM provided auditable, versioned outputs for enhanced compliance and reduced subjective errors.

Conclusions:

• KT-LLM offers a practical solution for auditable, clock-synchronized kidney transplantation workflows.
• The system enhances reproducibility, fairness monitoring, and standardization of clinical practice.
• KT-LLM addresses unmet needs for scalable and reliable kidney transplant care.