OHCA Classifier V11: Temporal + Location-Aware Model

Model Description

A transformer-based deep learning model for automatically identifying Out-of-Hospital Cardiac Arrest (OHCA) cases from clinical notes.

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

# 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

@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
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support