OHCA Classifier: Automated Out-of-Hospital Cardiac Arrest Identification

A transformer-based deep learning system for automatically identifying Out-of-Hospital Cardiac Arrest (OHCA) cases from clinical notes, with specific focus on reducing false positives from in-hospital cardiac arrest (IHCA) cases.

🎯 Key Results

Metric	V9 (Baseline)	V10 (+ Location)	V11 (+ Temporal)
Sensitivity	96.1%	84.2%	92.1%
Specificity	69.6%	89.6%	89.4%
F1-Score	0.732	0.814	0.856
AUC-ROC	0.932	0.938	0.956
IHCA False Positives	88 (64.2%)	39 (28.5%)	39 (28.5%)

Validation: 647 manually annotated clinical notes from UChicago C19 dataset

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/monajm36/ohca-classifier.git
cd ohca-classifier

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Using the Pre-trained Model

from predict_v11 import predict_ohca

# Example clinical note
note = """
Patient found unresponsive at home by family. 911 called.
EMS arrived and initiated CPR. ROSC achieved in field.
Transported to ED.
"""

# Predict
result = predict_ohca(note, threshold=0.14)

print(f"Prediction: {result['prediction']}")  # 'OHCA' or 'Non-OHCA'
print(f"Probability: {result['probability']:.2%}")
print(f"Features: {result['features']}")

Download Pre-trained Model

# Model is available on Hugging Face
# https://huggingface.co/monajm36/ohca-classifier-v11

# Or use the download script
python scripts/download_model.py

📊 Model Architecture

V11: Temporal + Location-Aware OHCA Classifier

Input Clinical Note
       ↓
┌──────────────────────┐
│  Text Processing     │
│  - Extract sections  │
│  - Tokenize (512)    │
└──────────────────────┘
       ↓
┌──────────────────────────────────────┐
│  Feature Extraction                  │
├──────────────────────────────────────┤
│  1. BERT Embeddings (768)            │
│  2. Location Features (2)            │
│     • OHCA indicators (22 phrases)   │
│     • IHCA indicators (25 phrases)   │
│  3. Temporal Features (5)            │
│     • Arrest timing score            │
│     • First location (in/out)        │
│     • Movement patterns              │
└──────────────────────────────────────┘
       ↓
┌──────────────────────┐
│  MLP Classifier      │
│  775 → 512 → 256 → 2 │
└──────────────────────┘
       ↓
  OHCA Probability

Base Model: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract

📁 Repository Structure

ohca-classifier/
├── README.md                          # This file
├── requirements.txt                   # Python dependencies
├── setup.py                          # Package installation
│
├── models/                           # Model code
│   ├── __init__.py
│   ├── v9_bert_classifier.py        # V9: Baseline BERT
│   ├── v10_location_aware.py        # V10: + Location features
│   └── v11_temporal_location.py     # V11: + Temporal features
│
├── training/                         # Training scripts
│   ├── train_v9.py
│   ├── train_v10.py
│   └── train_v11.py
│
├── prediction/                       # Prediction scripts
│   ├── predict_v9.py
│   ├── predict_v10.py
│   └── predict_v11.py
│
├── features/                         # Feature extraction
│   ├── __init__.py
│   ├── text_processing.py           # Section extraction
│   ├── location_features.py         # Location indicators
│   └── temporal_features.py         # Temporal features
│
├── evaluation/                       # Evaluation scripts
│   ├── compare_models.py            # V9 vs V10 vs V11
│   ├── threshold_optimization.py   # Find optimal thresholds
│   └── error_analysis.py           # Analyze false positives
│
├── scripts/                          # Utility scripts
│   ├── download_model.py            # Download from HF
│   └── prepare_data.py              # Data preprocessing
│
├── notebooks/                        # Jupyter notebooks
│   ├── model_comparison.ipynb
│   ├── feature_analysis.ipynb
│   └── demo.ipynb
│
├── docs/                            # Documentation
│   ├── COMPREHENSIVE_REPORT.pdf     # Full technical report
│   ├── model_architecture.md
│   ├── feature_engineering.md
│   └── training_guide.md
│
└── tests/                           # Unit tests
    ├── test_features.py
    ├── test_models.py
    └── test_predictions.py

🔬 Methodology

Progressive Feature Engineering

Our approach built three successive models, each addressing specific challenges:

V9: Baseline BERT Classifier

• Pure semantic understanding
• Issue: 65% of false positives were IHCA cases
• Learning: Model confused arrest terminology regardless of location

V10: Location-Aware Classifier

• Added 2 location features (22 OHCA + 25 IHCA indicators)
• Reduced IHCA false positives by 55.7%
• Issue: Lost sensitivity (96.1% → 84.2%)
• Learning: Location helps but temporal context missing

V11: Temporal + Location-Aware Classifier

• Added 5 temporal features (timing, movement patterns)
• Recovered sensitivity (84.2% → 92.1%)
• Maintained high specificity (89.4%)
• Learning: Temporal sequence crucial for disambiguation

Feature Categories

Location Features (2):

1. OHCA indicator count: home, EMS, scene, field, bystander, ambulance...
2. IHCA indicator count: floor, ICU, ward, room, bed, code blue...

Temporal Features (5):

1. Arrest timing score: "before arrival" vs "during hospitalization" phrases
2. First location outside: Binary indicator of first location mentioned
3. First location inside: Binary indicator of first location mentioned
4. Movement outside→inside: Count of transition patterns
5. Movement inside→inside: Count of transition patterns

📈 Training

Data

• Training Set: MIMIC-III clinical notes

  • 330 notes total (47 OHCA, 283 Non-OHCA)
  • Split: 70% train / 15% validation / 15% test
  • Average note length: 13,042 characters

• Validation Set: UChicago C19 dataset

  • 647 manually annotated cases
  • 203 OHCA (31.4%)
  • 137 IHCA (21.2%)
  • 307 Non-arrest (47.4%)

Train Your Own Model

# Train V11 model
python training/train_v11.py \
    --data_path /path/to/mimic_labelled_binary.csv \
    --output_dir ./models/v11_output \
    --batch_size 4 \
    --learning_rate 2e-5 \
    --num_epochs 5

# Train with custom hyperparameters
python training/train_v11.py \
    --config configs/custom_config.json

🎯 Threshold Selection

V11 offers flexible threshold tuning for different clinical scenarios:

Use Case	Threshold	Sensitivity	Specificity	F1	When to Use
Screening	0.14	92.1%	89.4%	0.856	Maximize recall
Balanced	0.74	82.3%	93.2%	0.831	General use
Research	0.85	75.4%	95.0%	0.810	High precision needed

# Use different thresholds
result_screening = predict_ohca(note, threshold=0.14)  # High sensitivity
result_balanced = predict_ohca(note, threshold=0.74)   # Balanced
result_research = predict_ohca(note, threshold=0.85)   # High specificity

📊 Evaluation

Compare All Models

# Run comprehensive comparison
python evaluation/compare_models.py \
    --data_path /path/to/c19_validation.csv \
    --output_dir ./results

# Generate comparison plots
python evaluation/compare_models.py --plot

Error Analysis

# Analyze false positives and false negatives
python evaluation/error_analysis.py \
    --model_path models/v11_output/final_model \
    --data_path /path/to/validation.csv

🔍 Example Predictions

True OHCA Case

Patient found unresponsive at home. Family called 911.
EMS arrived, started CPR. ROSC in field.
Transported to ED.

→ Prediction: OHCA (98.5%)
→ Key features:
   - Location: home (OHCA +1), EMS (+1)
   - Temporal: "found at home", "ROSC in field"
   - Movement: outside→inside

True IHCA Case

Patient admitted to medical floor for pneumonia.
On hospital day 3, found unresponsive in bed.
Code blue called. CPR initiated on floor.

→ Prediction: Non-OHCA (2.3% OHCA probability)
→ Key features:
   - Location: floor (+1), bed (+1)
   - Temporal: "hospital day 3", "admitted"
   - Movement: inside→inside

Challenging Case (Mixed Signals)

Patient arrested at home. EMS called, ROSC achieved.
Admitted to ICU. On hospital day 2, arrested again.

→ V9: OHCA (85%) - sees "home" and "EMS"
→ V10: Uncertain (52%) - conflicting locations
→ V11: OHCA (78%) - temporal features indicate primary arrest was OHCA

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Areas for improvement:

• Additional temporal features
• Multi-institution validation
• Support for non-English notes
• Real-time deployment pipeline
• Explainability visualizations

📝 Citation

If you use this code or model in your research, please cite:

@misc{moukaddem2025ohca,
  author = {Moukaddem, Mona},
  title = {OHCA Classifier: Automated Out-of-Hospital Cardiac Arrest Identification 
           using Temporal and Location-Aware Deep Learning},
  year = {2025},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/monajm36/ohca-classifier}},
  note = {Model available at \url{https://huggingface.co/monajm36/ohca-classifier-v11}}
}

📄 License

This project is licensed under the MIT License - see LICENSE file for details.

🙏 Acknowledgments

• MIMIC-III Database: Johnson et al., Scientific Data (2016)
• UChicago C19 Dataset: Validation data source
• BiomedNLP-PubMedBERT: Microsoft Research
• Hugging Face: Model hosting and transformers library

📧 Contact

Mona Moukaddem

• GitHub: @monajm36
• Hugging Face: @monajm36
• Email: [your-email@example.com]

🔗 Links

• 🤗 Pre-trained Model (Hugging Face)
• 📊 Interactive Demo (coming soon)
• 📄 Technical Report
• 📝 Blog Post (coming soon)

---

Built with ❤️ for improving cardiac arrest research and patient care