Google MedGemma Impact Challenge 2026 | Main Track + Novel Task Prize
A physician sees 20 patients a day. For each encounter, they spend an average of 16 minutes on documentation — 40% of every patient interaction lost to typing, clicking, and formatting. Across a career, that is roughly 24,000 hours spent on paperwork instead of patient care.
AI documentation tools were supposed to fix this. They didn't. They replaced one problem with another: instead of writing notes from scratch, physicians now spend their time editing AI-generated notes that are too long, too formal, and clinically imprecise. A 45-second encounter produces a 200-word wall of textbook prose that the physician must pare down to the 20 words that actually matter.
A practicing nephrologist who uses Dragon Copilot daily:
"More often than not I have to go and edit the notes and shorten them, because they read like textbook lexicon rather than shorthand designed to deliver efficient summaries with alacrity."
MedScribe was built to eliminate that editing step entirely.
MedScribe generates SOAP notes in the same concise shorthand that experienced clinicians write themselves. Not textbook prose. Not verbose summaries. Clinical shorthand — abbreviations, focused assessments, actionable plans.
| What existing tools produce | What MedScribe produces |
|---|---|
| "The patient is a 54-year-old female who presents to the emergency department with a chief complaint of shortness of breath." | "54 yo F c/o SOB." |
| "Assessment: After careful review of the imaging findings and clinical presentation, the most likely diagnosis is acute pulmonary embolism involving the segmental branches of the right lower lobe." | "Acute segmental PE, right lower lobe." |
| "Plan: It is recommended to initiate anticoagulation therapy. The patient's respiratory status should be closely monitored." | "1. Initiate anticoagulation. 2. Monitor resp status and O2 sat. 3. Pulm follow-up." |
A 200-word note becomes 104 words. No clinical information is lost. The physician reads it, confirms it, and moves on.
Current AI scribes stop at transcription. MedScribe continues. After generating the SOAP note, six clinical intelligence tools transform the static note into an actionable artifact:
- Billable Diagnoses — extracts diagnoses supported by documentation for ICD-10 coding, with evidence mapping
- Patient Summary — plain-language visit summary for patient portals
- Completeness Check — flags documentation gaps before the note is signed
- Differential Diagnosis — ranked DDx with supporting evidence
- Medication Check — interaction and contraindication screening
- Patient Intake Analysis — risk stratification, red flags, and screening recommendations from structured intake data
Each tool takes 10-18 seconds and operates on the generated SOAP note. The clinician runs only the tools they need.
MedScribe runs fully offline on consumer hardware. Patient data never leaves the device. No cloud API calls. No data logging. No third-party access. In healthcare, this isn't a feature — it's a requirement.
| Metric | Before MedScribe | With MedScribe | Savings |
|---|---|---|---|
| Documentation time | 16 min | ~5 min (review + confirm) | 11 min |
| Note editing after AI generation | 3-5 min | 0 min (note is chart-ready) | 3-5 min |
| ICD-10 code lookup | 2-3 min | 10 sec (auto-extracted) | ~2.5 min |
| Completeness check (mental) | 1-2 min | 15 sec (automated) | ~1.5 min |
| Patient summary for portal | 2-3 min | 18 sec (auto-generated) | ~2.5 min |
| Metric | Value |
|---|---|
| Encounters per year | 5,000 |
| Documentation time saved per encounter | ~11 min |
| Annual documentation time saved | ~917 hours |
| Additional patient encounters possible | ~1,100 (at 50 min/encounter) |
| Reduced after-hours charting ("pajama time") | Estimated 2-3 hrs/week |
These are projections based on published documentation burden data (Sinsky et al., Annals of Internal Medicine, 2016) and the measured difference between MedScribe's chart-ready notes and the editing time required for verbose AI output. Actual impact will vary by specialty, EHR system, and practice setting.
| Metric | Value |
|---|---|
| Quality score | 90/100 |
| Section completeness | 100% (S/O/A/P always present) |
| Hallucinated findings | 0% |
| WNL shortcuts | 0% (every finding explicitly stated) |
| Average word count | 104 words (vs 200+ for verbose tools) |
| Average SOAP inference | ~25s (RTX 5070 Ti) |
| Anti-hallucination | "Not documented in source" for missing data |
Three models from Google's HAI-DEF ecosystem in a single pipeline:
| Component | Model | Role |
|---|---|---|
| Speech Recognition | MedASR (105M, Conformer) | Medical dictation → text |
| SOAP Generation | MedGemma 1.5 4B + LoRA | Transcript → concise SOAP note |
| Clinical Intelligence | MedGemma 1.5 4B (base) | 6 post-generation analysis tools |
MedASR and MedGemma coexist in 16GB VRAM (4-bit quantization). The LoRA adapter adds 4.2M trainable parameters (0.1% of base) and was fine-tuned on 712 curated samples with anti-hallucination constraints. Total training cost: $1.28 via GPT-4o API.
For complete technical details — stopping criteria, CTC decoding pipeline, generation parameters, training configuration, and UI implementation — see TECHNICAL.md
Record or upload audio. MedASR transcribes, MedGemma converts to structured SOAP. The transcript is editable before generation — clinicians can correct transcription errors before the note is produced.
Paste any medical encounter transcript. Five example transcripts included covering cardiology, endocrinology, pediatrics, nephrology, and psychiatry.
Five tools operating on any generated or pasted SOAP note:
| Tool | What It Does | Time |
|---|---|---|
| Billable Diagnoses | Extracts ICD-10-codeable diagnoses with evidence | ~10s |
| Patient Summary | Plain-language summary for patients | ~18s |
| Completeness Check | Documentation gap detection | ~15s |
| Differential Diagnosis | Ranked DDx with evidence | ~18s |
| Medication Check | Interaction/contraindication screening | ~18s |
Structured form accepting demographics, history, medications, lifestyle, and labs. MedGemma produces risk assessments, differential considerations, recommended screenings, red flags, and clinical questions. Pre-filled with a demonstration case (58-year-old South Asian male with rapid-onset diabetes and declining renal function).
Side-by-side base MedGemma vs fine-tuned output on the same transcript. Demonstrates the concrete value of LoRA fine-tuning: 46% shorter notes, clinical shorthand, zero hallucination, 100% section completeness.
- Python 3.10+
- CUDA 12.1+ compatible GPU (8GB minimum, 16GB recommended)
- HuggingFace access to
google/medgemma-4b-itandgoogle/medasr
git clone https://github.com/Tushar-9802/MedScribe-1.git
cd MedScribe-1
python -m venv venv
source venv/bin/activate # Linux/Mac
venv\Scripts\activate # Windows
pip install -r requirements.txtpython app.pyOpens at http://localhost:7860. Click Load Models to load MedASR and
MedGemma into VRAM, then navigate to any tab.
- English only — MedASR training constraint
- Research prototype — not validated for clinical use
- Synthetic training data — 712 samples, not real clinical encounters
- Hardware dependent — ~25s inference on RTX 5070 Ti
- Clinical tools not fine-tuned — use base model instruction-following
MedScribe-1/
app.py # Gradio UI (main entry point)
requirements.txt # Runtime dependencies
README.md # This file
TECHNICAL.md # Architecture and implementation deep-dive
src/
__init__.py
pipeline.py # MedASR + MedGemma orchestration
inference.py # Model loading, generation, stopping criteria
models/checkpoints/
medgemma_v2_soap/final_model/ # LoRA adapter weights
data/processed/ # Training data
train_v2.py # LoRA fine-tuning script
evaluate_v2.py # Quality evaluation
generate_soap_gpt4o.py # Training data generation
MIT License (code). Model weights subject to Google's terms of use.
Adaptors/Safetensors: Huggingface
Dataset: Kaggle
Technical Documentation: TECHNICAL.md
GitHub: @Tushar-9802