🔬 Medical AI System Documentation
Project Overview
This project contains two advanced AI systems for medical imaging and risk assessment:
- Pregnancy Risk Prediction Model - Predicts pregnancy complications using clinical data
- Fetal Ultrasound Plane Classification - Classifies fetal ultrasound images into anatomical planes
🤱 Pregnancy Risk Prediction Model
Model Performance
- Algorithm: Random Forest Classifier
- Accuracy: 100% on test data
- Features: 11 clinical parameters
- Classes: High Risk, Low Risk
- Dataset: 1,187 patient records
Key Features
- Age, Blood Pressure (Systolic/Diastolic)
- Blood Sugar, Body Temperature, BMI
- Medical History (Previous Complications, Diabetes)
- Mental Health, Heart Rate
Feature Importance (Top 5)
- Blood Sugar (BS): 22.8%
- Preexisting Diabetes: 21.6%
- Heart Rate: 16.0%
- BMI: 14.7%
- Gestational Diabetes: 8.5%
Model Metrics
Classification Report:
precision recall f1-score support
High 1.00 1.00 1.00 95
Low 1.00 1.00 1.00 143
accuracy 1.00 238
macro avg 1.00 1.00 1.00 238
weighted avg 1.00 1.00 1.00 238
🔬 Fetal Ultrasound Plane Classification
Model Performance
- Algorithm: Vision Transformer (ViT-Base-Patch16-224)
- Validation Accuracy: 91.69%
- Training Time: 18.5 minutes (Apple Silicon M4)
- Dataset: 12,400 ultrasound images
- Classes: 9 anatomical plane categories
Training Configuration
- Device: Apple Silicon MPS (Metal Performance Shaders)
- Batch Size: 2 (thermal-optimized)
- Epochs: 2
- Learning Rate: 5e-5
- Architecture: ARM64 optimized
Classification Categories
Fetal Brain Planes (4 types)
- Trans-thalamic: 1,638 images
- Trans-cerebellum: 714 images
- Trans-ventricular: 597 images
- Other brain views: 143 images
Anatomical Structures (4 types)
- Fetal thorax: 1,718 images
- Maternal cervix: 1,626 images
- Fetal femur: 1,040 images
- Fetal abdomen: 711 images
Quality Control (1 type)
- Other/Unclear: 4,213 images
Training Metrics
Final Training Loss: 0.21
Validation Loss: 0.316
Training Speed: 4.47 iterations/second
System Resources:
- CPU Usage: 5.4% (post-training)
- Memory Usage: 65.3%
- Temperature: Stable (no overheating)
Apple Silicon Optimizations
- MPS Acceleration: Full M4 chip utilization
- Thermal Management: Prevented overheating
- Memory Efficiency: Optimized batch sizes
- Native Performance: ARM64 PyTorch builds
🏗️ System Architecture
Project Structure
hackathon15092025/
├── src/ # Source code
│ ├── app.py # Pregnancy risk Streamlit app
│ └── pregnancy_risk_prediction.py
├── fetal_plane_app.py # Fetal plane Streamlit app
├── fetal_plane_classifier.py # Training script
├── models/ # Trained models
│ ├── pregnancy_risk_model.pkl
│ └── fetal_plane_model/
├── data/ # Datasets
│ └── Dataset - Updated.csv
├── FETAL_PLANES_ZENODO/ # Ultrasound dataset
├── static/css/ # Styling
├── index.html # Main dashboard
└── requirements*.txt # Dependencies
Technology Stack
- Machine Learning: scikit-learn, PyTorch, Transformers
- Web Framework: Streamlit
- Frontend: HTML5, CSS3, JavaScript
- Visualization: Plotly, Matplotlib
- Deployment: Apple Silicon optimized
🚀 Deployment Guide
Prerequisites
- Python 3.9+
- macOS with Apple Silicon (M1/M2/M3/M4)
- 8GB+ RAM recommended
Installation
cd /Users/karthik/Projects/hackathon15092025
pip install -r requirements.txt
pip install -r requirements_fetal.txt
python src/pregnancy_risk_prediction.py
python train_fetal_model_thermal.py
Running Applications
streamlit run src/app.py
streamlit run fetal_plane_app.py --server.port 8502
open index.html
📊 Performance Benchmarks
Pregnancy Risk Model
| Metric |
Value |
| Training Accuracy |
100% |
| Validation Accuracy |
100% |
| Inference Time |
<1ms |
| Model Size |
2.3MB |
| Features |
11 |
Fetal Plane Model
| Metric |
Value |
| Training Accuracy |
95.4% |
| Validation Accuracy |
91.69% |
| Inference Time |
<100ms |
| Model Size |
346MB |
| Parameters |
86M |
System Performance (M4 MacBook)
| Resource |
Usage |
| CPU |
5.4% (idle) |
| Memory |
65.3% |
| GPU (MPS) |
Active |
| Temperature |
Stable |
🔒 Security & Privacy
Data Protection
- No Data Storage: Patient data not permanently stored
- Local Processing: All inference runs locally
- HIPAA Considerations: Designed for privacy compliance
- Secure Models: No data leakage in model weights
Recommendations
- Use in controlled medical environments
- Implement proper access controls
- Regular security audits
- Compliance with local regulations
🎯 Clinical Applications
Pregnancy Risk Assessment
- Primary Care: Initial risk screening
- Obstetrics: Prenatal care planning
- Emergency: Rapid risk evaluation
- Telemedicine: Remote consultations
Ultrasound Classification
- Radiology: Image quality control
- Training: Medical education tool
- Workflow: Automated image sorting
- Research: Large-scale studies
⚠️ Limitations & Disclaimers
Model Limitations
- Educational Purpose: Not for clinical diagnosis
- Validation Needed: Requires clinical validation
- Population Bias: Trained on specific datasets
- Continuous Learning: Models need regular updates
Usage Guidelines
- Always consult qualified healthcare professionals
- Use as decision support, not replacement
- Validate results with clinical judgment
- Report unusual predictions for review
📈 Future Enhancements
Planned Features
- Multi-language Support: International deployment
- Real-time Monitoring: Continuous risk assessment
- Integration APIs: EHR system connectivity
- Advanced Models: Transformer-based improvements
Research Directions
- Federated Learning: Multi-site model training
- Explainable AI: Enhanced interpretability
- Edge Deployment: Mobile device optimization
- Clinical Trials: Prospective validation studies
📞 Support & Contact
Technical Support
- Documentation: This file and README files
- Issues: Check terminal logs for errors
- Performance: Monitor system resources
- Updates: Regular dependency updates
Development Team
- AI/ML Engineering: Model development and optimization
- Medical Informatics: Clinical workflow integration
- Software Engineering: Application development
- Quality Assurance: Testing and validation
Last Updated: January 2025
Version: 1.0
Platform: Apple Silicon Optimized
