Mic52/tabnet-tachycardia-mimiciv

TabNet Deep Model for Tachycardia Detection (MIMIC-IV ECG)

This repository hosts a TabNet deep learning model trained to classify episodes of tachycardia based on ECG-derived measurements from the MIMIC-IV ECG dataset.

🔍 Problem

Tachycardia is defined here as a RR interval < 600ms, derived directly from the rr_interval column. The goal is to predict the presence of tachycardia using temporal, positional and axis-based ECG features.

📊 Dataset

• Original Source: MIMIC-IV ECG (v1.0)
• Processed via pandas and sklearn (see notebook).
• Train/Val/Test split:
  • Train: 560,003 samples
  • Val: 120,026 samples
  • Test: 120,006 samples
• Tachycardia distribution:
  ~17% Positive class (imbalanced)

🧠 Model

• Architecture: PyTorch TabNet
• Type: Deep Tabular Neural Network
• Framework: PyTorch
• Training details:
  • Early Stopping: 27 epochs
  • Best Validation Accuracy: 98.67%
  • Final Test Accuracy: 99.0%
  • Precision (tachycardia): 1.00
  • Recall (tachycardia): 0.90
  • F1 Score (tachycardia): 0.95

🧪 Results

Metric	Value
Accuracy	99.0%
F1 Score	0.95
Precision	1.00
Recall	0.90

📁 Files Included

• tabnet_deep_model.zip → Trained TabNet model
• MIMIC_ECG_Tachycardia_Analysis.ipynb → Full preprocessing + metrics

🛠 How to Use

from pytorch_tabnet.tab_model import TabNetClassifier

model = TabNetClassifier()
model.load_model("tabnet_deep_model.zip")
preds = model.predict(X_test)

⚠️ License & Data Use

This model was trained on the MIMIC-IV ECG dataset, which requires credentialed access via PhysioNet. Please ensure compliance with PhysioNet's data use policy.

---

Trained and released by Mic52.