# A Large Language Model Workflow for Auditable Brain Abscess Risk Stratification and Pre-residency Scholarship: A Technical Report

## TL;DR

This paper introduces a structured workflow using large language models to teach medical students how to build auditable AI models for brain abscess risk stratification.

## Contribution

A novel educational framework using LLMs with a prompt ledger to teach rigorous, transparent AI use in medical training.

## Key findings

- A Neurologic Deterioration in Brain Abscess Score (NDBAS v0.1) was developed and applied to a clinical case.
- The LLM-assisted workflow accelerated evidence review while maintaining rigor through verification and oversight.
- Three curriculum artifacts were produced: a narrative case appendix, a prompt ledger, and a variable dictionary.

## Abstract

Large language models (LLMs) are increasingly available to medical trainees, but transparent and auditable methods for incorporating them into scholarly work and emerging artificial intelligence (AI) literacy curricula remain limited. This technical report describes a mentored educational framework in which a de-identified brain abscess case report was transformed into a reproducible, LLM-supported risk-stratification model to teach rigor, verification, and structured AI use to a fourth-year medical student. A single clinical case was reconstructed using structured variables from history, imaging, laboratory trends, and symptom trajectory. Each interaction with the LLM followed a standardized query pattern and was logged in a prompt ledger capturing prompts, rationales, and inclusion or exclusion decisions; chain-of-thought outputs were retained only as reasoning traces for supervised review. Literature-supported deterioration indicators were synthesized into a provisional Neurologic Deterioration in Brain Abscess Score (NDBAS v0.1), emphasizing lesion diameter, midline shift, vasogenic edema, inflammatory markers, sensorium changes, and metabolic risk factors. Applying NDBAS v0.1 to the index case yielded a score of 5, corresponding to a moderate-risk tier. Relative to the learner’s prior manual approach, the LLM-assisted synthesis accelerated evidence review while maintaining rigor through primary-source verification and faculty oversight and produced three concrete curriculum artifacts: a narrative case appendix, a prompt ledger, and a variable dictionary with domain-specific cut points and tiered risk thresholds. This structured, audit-ready LLM workflow offers a practical route to building pre-residency AI literacy by embedding a supervised “Prompt Ledger plus ask-verify-revise” pattern into complex case-based learning, moving students from casual chatbot use to transparent evidence synthesis, critical appraisal, and conceptual model-building from single cases. With appropriate governance, de-identification, and faculty oversight, similar mentored workflows can help learners at other institutions turn single cases into auditable conceptual models, strengthening both their scholarly output and their readiness to engage with AI responsibly during clinical training.