Key Innovation: Combines semantic understanding (PubMedBERT) with explicit location and temporal features to distinguish OHCA from in-hospital cardiac arrest (IHCA).

Training Data

Dataset: MIMIC-III clinical notes
Size: 330 notes (47 OHCA, 283 Non-OHCA)
Split: 70% train / 15% validation / 15% test
Average note length: 13,042 characters

Performance (C19 Validation - 647 notes)

Metric	Score
Sensitivity	92.1%
Specificity	89.4%
Precision	79.9%
F1-Score	0.856
AUC-ROC	0.956

Model Architecture

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

Input Features (775 dimensions):

BERT embeddings: 768
Location features: 2
- OHCA location indicator count (22 phrases)
- IHCA location indicator count (25 phrases)
Temporal features: 5
- Arrest timing score (when arrest occurred)
- First location outside hospital (binary)
- First location inside hospital (binary)
- Movement outside→inside count
- Movement inside→inside count

Classifier: 3-layer MLP (775 → 512 → 256 → 2)

Key Features

Location Features

OHCA indicators: home, EMS, scene, field, bystander, ambulance, paramedics, etc.

IHCA indicators: floor, ICU, ward, room, bed, code blue, admitted, telemetry, etc.

Temporal Features

Captures the story of what happened:

When: Before arrival vs during hospitalization
Where it started: First location mentioned (inside/outside)
How patient moved: Direction of transitions (outside→inside vs inside→inside)

Usage

python

# Note: Requires custom model class and feature extraction
# See model files for implementation details

from transformers import AutoTokenizer
import torch

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("monajm36/ohca-classifier-v11")

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

# Extract features (requires custom code)
# location_features = extract_location_features(note)
# temporal_features = extract_temporal_features(note)

# Tokenize
inputs = tokenizer(note, return_tensors="pt", max_length=512, truncation=True)

# Predict (requires loading custom model architecture)
# ...

Threshold Selection

Choose threshold based on your clinical use case:

Use Case	Threshold	Sensitivity	Specificity	F1
Screening (High Recall)	0.14	92.1%	89.4%	0.856
Balanced	0.74	82.3%	93.2%	0.831
Research (High Precision)	0.85	75.4%	95.0%	0.810

Limitations

Trained on single institution (MIMIC-III)
May not generalize to all clinical documentation styles
IHCA false positive rate: ~28.5% at optimal threshold
Requires feature extraction code (not included in model weights)
Best performance on notes with clear EMS or location context

Model Versions

This is Version 11 - the latest and most accurate version.

Version	Key Features	F1-Score
V9	BERT only	0.732
V10	+ Location features	0.814
V11	+ Temporal features	0.856

Citation

bibtex

@misc{moukaddem2025ohca,
  author = {Moukaddem, Mona},
  title = {OHCA Classifier V11: Temporal and Location-Aware Model for Out-of-Hospital Cardiac Arrest Identification},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/monajm36/ohca-classifier-v11}}
}

Contact

For questions, issues, or collaboration opportunities, please open an issue on the model repository.

Model Card Authors

Mona Moukaddem

Acknowledgments

Training data: MIMIC-III Clinical Database
Validation data: UChicago C19 dataset
Base model: Microsoft BiomedNLP-PubMedBERT