MedFill — AI-Powered Medical Report Completion
Upload any partial blood test report → Get a complete 25-biomarker lab panel in ~15 seconds.
Local-first, offline, zero cloud. Your patient data never leaves your machine.
Real-world medical reports are incomplete . Tests are marked PENDING, values are missing due to cost or urgency, and doctors are forced to make decisions with partial data. MedFill fills those gaps instantly using AI.
📷 Partial blood report image (Apollo, Thyrocare, any lab)
↓
EasyOCR (GPU, ~5s) — reads every word on the report
↓
Regex Parser — extracts 12 known biomarker values
↓
TabularImputerModel — predicts all 13 missing values
(Transformer · PyTorch) R² = 99.7%
↓
✅ Complete 25-Feature Lab Panel
Hemoglobin : 10.8 g/dL [extracted]
MCH : 27.3 pg [extracted]
MCV : 90.8 fL [★ AI predicted]
Creatinine : 4.6 mg/dL [★ AI predicted]
WBC Count : 9,166 /µL [★ AI predicted]
...
Metric
Value
R² Score 0.9968 (99.7%)
Validation Loss (MSE) 0.001632
MAE (normalized) 0.0205
Training epochs 76
Val patients 2,231
Training patients 12,637
OCR Confidence 0.81 (81%)
End-to-end time ~15 seconds
┌──────────────────────────────────────────────────────────────────┐
│ React Frontend (localhost:5173) │
│ FileUpload → PatientDashboard → Report Analysis Panel │
└──────────────────────────┬───────────────────────────────────────┘
│ POST /api/v1/analyze
┌──────────────────────────▼───────────────────────────────────────┐
│ FastAPI Backend (localhost:8000) │
│ │
│ ① EasyOCRAgent GPU OCR → raw text (~5s, conf=0.81) │
│ ② direct_parse_ocr() Regex → 12 extracted features │
│ ③ impute() Transformer → 13 predicted features │
│ ④ Response builder patient + lab_panels + complete_panel │
└──────────────────────────┬───────────────────────────────────────┘
│
┌──────────────────────────▼───────────────────────────────────────┐
│ TabularImputerModel (ml_pipeline/) │
│ │
│ Input: 25 features (missing → 0) │
│ Embed: Linear(1→128) + Positional Embedding(25,128) │
│ Encoder: 4× TransformerEncoderLayer │
│ (8 heads · dim=128 · FFN=512 · Pre-LN · dropout=0.1) │
│ Output: Linear(128→1) per feature → 25 reconstructed values │
│ Params: 796,673 │
└──────────────────────────────────────────────────────────────────┘
Dataset
Source
Patients
Features
Anemia Dataset Kaggle (CC0) 1,421
Hemoglobin, MCH, MCHC, MCV, Gender
CKD Dataset Kaggle — CKD Disease · donated by Apollo Hospitals400
25 biomarkers (renal, electrolytes, urinalysis)
NHANES CDC NHANES 10,816
Blood count, metabolic, renal panels
Combined —
12,637 25 unified features
1. Install Python dependencies pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt 2. Train the imputation model python -m ml_pipeline.train
# ~30s on GPU · saves ml_pipeline/checkpoints/best_model.pt 3. Run inference from CLI python infer.py path/to/report.jpg uvicorn api_gateway.main:app --reload --port 8000
# Swagger docs: http://localhost:8000/docs 5. Start the React frontend cd Fronted_medfill
npm install
npm run dev
# Open: http://localhost:5173
Method
Endpoint
Description
POST/api/v1/analyzeUpload report image → complete lab panel (OCR + imputation in one call)
GET/healthBackend status check
GET/docsSwagger UI
Example:
curl -X POST http://localhost:8000/api/v1/analyze \
-F " file=@report.jpg" Response includes:
{
"patient" : { "name" : " MR. RAHUL SHARMA" "age_years" : 52 , "sex" : " Male" "lab_panels" : [
{
"panel_name" : " Complete Blood Count (CBC)" "results" : [
{ "test_name" : " Hemoglobin (Hb)" "value" : 10.8 , "flag" : " low" "unit" : " g/dL" "test_name" : " MCH" "value" : 27.3 , "flag" : " normal" "unit" : " pg" "complete_panel" : {
"Hemoglobin" : { "value" : 10.8 , "source" : " extracted" "predicted" : false },
"MCV" : { "value" : 90.8 , "source" : " predicted" "predicted" : true }
},
"ocr_confidence" : 0.81 ,
"n_extracted" : 12 ,
"n_predicted" : 13 ,
"elapsed_s" : 14.7
}
Report Type
Result
CBC / Blood Count
✅ Best
Renal Function Test
✅ Great
Diabetes / Glucose panel
✅ Good
Apollo / Thyrocare / printed reports
✅ Full support
Blurry / poor lighting
⚠️ Partial (model predicts more)
Handwritten
⚠️ Reduced accuracy
Hindi / regional language
❌ English OCR only
Radiology / MRI
❌ No numeric biomarkers
Performance Comparison (OCR Engine)
Metric
Old (llava:7b VLM)
New (EasyOCR)
OCR time
110 seconds
5.5 seconds
Confidence
0.50
0.81
Hallucinations
Yes
No
VRAM
4.7 GB
~1 GB
Ollama required
Yes
No
report classifier/
├── ai_agents/
│ ├── easyocr_agent.py ← GPU OCR engine (primary)
│ ├── pipeline.py ← Orchestrator
│ ├── vision_agent.py ← Legacy VLM OCR
│ ├── structuring_agent.py ← llama3 backup (optional)
│ └── schemas.py ← Pydantic models
├── api_gateway/
│ └── main.py ← FastAPI (all endpoints)
├── Fronted_medfill/
│ ├── src/
│ │ ├── App.jsx
│ │ ├── api.js
│ │ └── components/
│ │ ├── FileUpload.jsx
│ │ ├── PatientDashboard.jsx
│ │ ├── PredictionPanel.jsx
│ │ └── Navbar.jsx
│ └── package.json
├── ml_pipeline/
│ ├── train.py ← Training loop
│ ├── data/dataset.py ← Data loading (NHANES + Anemia + CKD)
│ └── checkpoints/ ← best_model.pt (gitignored)
├── data/raw/ ← CSV datasets (gitignored)
├── infer.py ← CLI inference + impute()
├── requirements.txt ← Python dependencies
└── README.md
Component
Tested On
GPU
NVIDIA RTX 4050 Laptop (6 GB VRAM)
RAM
16 GB
Python
3.11
PyTorch
2.5.1 + CUDA 12.1
Name
Role
Pradyuman ML pipeline · Transformer model · FastAPI backend · EasyOCR integration
Kshitij Dataset research · Model evaluation · Clinical validation
Zulikha React frontend · UI/UX design · API integration
Parth Bansal Project analysis · Research · Presentation & pitch
Prince Project analysis · Research · Presentation & pitch
MIT