haguman/discharge-navigator-medgemma

Discharge Navigator — MedGemma 4B Application

Model Description

Discharge Navigator is a prompt-engineering application built on google/medgemma-4b-it, Google's HAI-DEF (Health AI Developer Foundations) model for clinical text understanding.

This is not a fine-tuned model. No custom weights are provided. This repository documents the prompt engineering methodology, evaluation results, and schema validation pipeline that transform MedGemma 4B into a clinical discharge note extraction system.

Base Model Lineage

Component	Details
Base model	google/medgemma-4b-it
Quantized variant	williamljx/medgemma-4b-it-Q4_K_M-GGUF (for edge/CPU)
Parameters	4B
Precision	bfloat16 (GPU) / Q4_K_M GGUF (CPU)
Fine-tuning	None — prompt engineering only

Intended Use

Extract structured discharge packets from clinical notes, producing:

• Diagnoses with candidate ICD-10 codes and confidence levels
• Medications with dosing details (dose, route, frequency)
• Follow-up instructions with urgency levels
• Red flags requiring immediate attention
• Missing information gaps with severity ratings
• Evidence spans grounding each claim to exact source text

Not a medical device. All outputs require clinician review before clinical use.

Prompt Engineering Approach

Dual-Variant Strategy

The system uses two prompt variants with a retry escalation strategy:

Attempt	Variant	Temperature	Strategy
1	A (contract-style)	0.0	Detailed role + rules + schema
2	A (contract-style)	0.1	Slight randomness for diversity
3	B (strict fallback)	0.0	Minimal prompt, brevity-focused

Key Prompt Design Principles

1. Schema-first extraction — prompt specifies exact JSON structure with all field names, types, and allowed enum values
2. Evidence grounding mandate — every extraction must include evidence_spans that are exact substrings of the source note (max 20 words, max 5 per item)
3. Confidence calibration — three-level confidence (low/medium/high) with explicit instructions to use "low" when uncertain
4. Negative instructions — "DO NOT GUESS. If it is not in the note, do not include it."

Schema Validation Gate

All model outputs pass through a Pydantic v2 schema (DischargePacket) that enforces:

• Required fields: diagnoses, medications, followups, red_flags, patient_summary, missing_info
• Enum constraints: confidence (low/medium/high), urgency (routine/soon/urgent)
• Evidence span caps: max 5 spans per item, max 20 words per span (adaptive token budget)
• Type coercion for common LLM output quirks (null strings, missing lists)

Invalid outputs are rejected and retried. No malformed data reaches consumers.

Evaluation Results

Evaluated on 50 clinical notes from MTSamples (CC0 license), CPU-only inference.

Metric	Value	Status
Parse rate (valid JSON)	92% (46/50)	PASS
Median latency (CPU)	34s	PASS
Diagnosis grounding	94.4%	PASS
Medication grounding	94.5%	PASS
Overall grounding	87.5%	PASS
Follow-up grounding	76.1%	IMPROVING

Failure Analysis

4/50 notes failed structured parsing:

• 3 token limit truncation (output exceeded 4096 tokens)
• 1 missing required field (missing_info omitted)

All failures were caught by the schema validation gate before reaching consumers.

Inference Backends

Backend	Environment	Model ID	Size
HuggingFace Transformers	Kaggle T4 GPU	google/medgemma-4b-it	~8 GB (bf16)
Ollama (GGUF)	Local CPU / Edge	williamljx/medgemma-4b-it-Q4_K_M-GGUF	2.5 GB

The GGUF quantization enables fully offline, air-gapped deployment on consumer hardware — no GPU, no internet required.

India Context

Designed for resource-constrained healthcare settings:

• 48M hospitalizations/year in India
• 1:11,000 doctor-to-patient ratio in rural areas
• Zero internet dependency enables rural deployment
• CPU-only inference runs on standard hospital workstations

Links

• Live Demo
• GitHub Repository
• MedGemma Impact Challenge

Citation

@misc{discharge-navigator-2026,
  title={Discharge Navigator: Offline Clinical Discharge Note Extraction with MedGemma},
  year={2026},
  url={https://github.com/LegenDairy93/discharge-navigator}
}