Medical AI Assistant for Battlefield Care Using LLMs

johns_hopkins · Healthcare · 2023 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Battlefield medical decision support

Untrained soldiers may need to provide prolonged trauma care in chaotic battlefield settings, but traditional clinical decision-support systems assume trained users, structured inputs, and calm clinical environments.

Constraints:

• The user may have no specialized medical knowledge.
• The environment is chaotic and operational rather than a calm clinical setting.
• Medical guidance must be accurate, understandable, and free of unnecessary jargon.
• Life-critical use makes hallucination, uncertainty, and opacity especially dangerous.
• The project was explicitly still a proof of concept and not battle-ready.

Architecture

Johns Hopkins APL built CPG-AI, a proof-of-concept conversational medical assistant on top of its RALF framework. RALF supplies LLM app-building tools and conversational-agent tooling; CPG-AI combines those with carefully engineered prompts, Tactical Combat Casualty Care flowchart algorithms translated into machine-readable form, and more than 30 Department of Defense clinical practice guidelines converted into text. The prototype infers patient condition from conversation, answers plain-English questions, and guides the user through tactical field care algorithms.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  User["Battlefield user"] --> Agent["CPG-AI conversational agent"]
  Agent --> RALF["RALF framework"]
  RALF --> LLM["Large language model"]
  Agent --> TCCC["TCCC care algorithms"]
  Agent --> Guidelines["Clinical practice guideline text"]
  LLM --> TCCC
  LLM --> Guidelines

Components

Data flow

A soldier describes a casualty in natural language. The CPG-AI conversational agent uses RALF to route that interaction through an LLM-backed application, injects situational prompts and ground-truth care algorithms, retrieves or categorizes relevant clinical practice guideline text, and returns plain-English guidance that can alternate between step-by-step care algorithms and ad-hoc questions.

Why this architecture

APL chose an LLM-based conversational architecture because the task is naturally language-heavy and previously would have required many separately trained components. Grounding the LLM in care algorithms and clinical guidelines addressed the high-stakes need for accurate medical guidance, while RALF made the implementation reusable across LLM applications and conversational agents.

Models used: large language model · ChatGPT-like model family

Scale

• The traditional architecture might require 20 or 30 separate behind-the-scenes components.
  ““There might be 20 or 30 individual components running behind the scenes to enable a conversational agent to help soldiers assist their buddies on the battlefield”article, stated
• The team converted more than 30 clinical practice guidelines for ingestion.
  “the researchers found and converted more than 30 clinical practice guidelines from the Department of Defense Joint Trauma System to be ingested as text by their model”article, stated
• The prototype covers common battlefield injuries including breathing issues, burns, and bleeding.
  “tactical field care — a category of care that encompasses the most common injuries encountered on the battlefield, including breathing issues, burns and bleeding.”article, stated

Alternatives considered (roads not taken)

Traditional structured clinical-support AI

Rejected because: It requires calibrated rules and labeled training data and is better suited to experts in calm settings than novice battlefield caregivers.

Many bespoke neural networks for search, relevance, and dialogue management

Rejected because: APL used a single LLM adapted through prompts and context instead of training separate task-specific models.

Deploy the proof of concept as battle-ready

Rejected because: The team explicitly said the prototype was only a step toward operational readiness.

Patterns

rag-retrieval (knowledge_grounding)

CPG-AI grounds model responses in Tactical Combat Casualty Care algorithms and converted DoD clinical practice guideline text.

context-engineering (context_management)

The team relies on precise prompts plus situational context and relevant medical information to adapt a general LLM to battlefield care.

tool-use-orchestration (tooling)

RALF separates LLM application tools from conversational agents that can take advantage of those applications.

safety-policy-enforcement (control)

The project frames accuracy, transparency, reasoning, uncertainty, and non-battle-ready status as safety constraints for high-stakes deployment.

Tradeoffs

Use one adaptable LLM instead of many specialized models

Chose A prompted LLM with injected medical ground truth over 20 or 30 task-specific components with bespoke neural networks.

Rationale: The LLM reduces the need for many trained components and can respond fluidly to natural-language situations, but it shifts difficulty into prompts and grounding.

Ground the model with formal care algorithms

Chose TCCC flowcharts and clinical practice guidelines as source material over Relying only on pretrained model knowledge.

Rationale: High-stakes medical guidance requires authoritative protocols, but those protocols have to be converted into model-usable form.

Keep CPG-AI in proof-of-concept phase

Chose Measured prototype development and future improvements over Claiming immediate battlefield deployment readiness.

Rationale: The project showed promising capabilities but still needed expanded coverage, better prompts, and more reliable guideline retrieval before operational use.

Gotchas (what bit them)

General LLM capability does not imply reliable battlefield medical behavior.

Root cause: LLMs can perform some tasks well and others badly, and those failure boundaries are not obvious in advance.

Resolution: The team combined careful prompt engineering with injected care-algorithm ground truth and kept the project at proof-of-concept status.

Structured clinical AI assumptions do not match novice battlefield use.

Root cause: Traditional systems rely on precise rules, labeled data, and expert users in calmer environments.

Resolution: APL used a conversational LLM agent intended to guide novices in plain English.

Prototype success does not settle operational readiness.

Root cause: The battlefield setting imposes high accuracy, transparency, uncertainty, latency, and practical deployment demands not solved by the first phase.

Resolution: The next phase focuses on expanding conditions, better prompts, and improved categorization/retrieval from guidelines.

Evaluation

The article reports first-phase prototype capabilities and planned next-phase improvements rather than a formal battlefield evaluation. Evidence centers on demonstrated behaviors: condition inference, accurate non-jargon answers, care-algorithm guidance, and switching between procedural guidance and questions.

Metrics: prototype status · ability to infer patient condition from conversational input · accurate answers without medical jargon · coverage of tactical field care injuries such as breathing issues, burns, and bleeding · more than 30 clinical practice guidelines converted

Outcomes

• APL produced a proof-of-concept CPG-AI prototype for battlefield medical guidance.
• The prototype can infer patient condition, answer questions without jargon, and guide users through tactical field care algorithms.
• RALF provided reusable LLM application and conversational-agent tooling.
• The team converted TCCC flowcharts and more than 30 DoD clinical practice guidelines into model-usable knowledge.
• The project remained explicitly not battle-ready and planned further work on prompts, condition coverage, and guideline retrieval.

The article reports a first-phase prototype, more than 30 converted guidelines, coverage of common tactical field care injuries, and future expansion rather than production deployment volume.