Data Extraction & Analysis

Sudden Cardiac Death Database (SDDB) and Normal Sinus Rhythm Database (NSDB)

This project utilizes the Sudden Cardiac Death Database (SDDB) provided by PhysioNet to extract and preprocess ECG data for research purposes. The focus is on patients who experienced ventricular fibrillation (VF). We specifically extract 20-second continuous rhythm segments of ECG data at different time intervals before the onset of VF.

Overview

Sudden Cardiac Death Database (SDDB)

The SDDB contains Holter recordings from patients who experienced sustained ventricular tachyarrhythmia and, in many cases, actual cardiac arrest. The database includes:

• 18 patients with underlying sinus rhythm
• 1 patient who was continuously paced
• 4 patients with atrial fibrillation

Normal Sinus Rhythm Database (NSDB)

The NSDB consists of ECG recordings from patients with normal sinus rhythm, providing a baseline for comparison with the ECG data from SDDB patients.

Data Extraction Process

Objectives

For each patient who experienced VF, the following ECG data segments are extracted:

• 20-second continuous ECG rhythm segments from 30 to 60 minutes before VF onset.
• 20-second continuous ECG rhythm segments from 60 to 120 minutes before VF onset.
• 20-second continuous ECG rhythm segments from 120 to 180 minutes before VF onset.

For patients in the NSDB, similar 20-second continuous ECG rhythm segments are extracted to serve as a baseline for comparison.

Data Extraction Procedure

1. Load Data:

   • The ECG data and annotation files are loaded using the WFDB Python package.

2. Convert Time:

   • VF onset times and durations are converted into seconds for precise indexing.

3. Segment Extraction:

   • For SDDB Patients:
     • For each patient, extract 20-second continuous ECG rhythm segments at the specified intervals:
       • 30 to 60 minutes before VF onset
       • 60 to 120 minutes before VF onset
       • 120 to 180 minutes before VF onset
   • For NSDB Patients:
     • Randomly extract 20-second continuous ECG rhythm segments from the available data to establish a baseline normal sinus rhythm.

4. Data Saving:

   • Extracted segments are saved as .npy files for further analysis.
   • For the NSDB, ensure that the segments are extracted from baseline periods of normal sinus rhythm.

5. Comparison

   • Perform a comparative analysis between the ECG segments from SDDB (pre-VF) and the baseline segments from the NSDB to identify patterns or anomalies associated with the onset of VF.

Directory Structure

• src/: Source code for data extraction and processing.
• preprocess/: Contains scripts for data extraction and preprocessing.
• models/: Contains scripts for modeling and evaluation.
• data/: Directory containing raw ECG data files.
• results/: Directory where processed .npy files are saved.

Example Usage

Run the provided Python scripts to perform data extraction as follows:

python src/preprocess/generate_ecg_segments_from_vf_onset.py --interval_start 30 --interval_end 60 
python src/preprocess/generate_ecg_segments_from_vf_onset.py --interval_start 60 --interval_end 120 
python src/preprocess/generate_ecg_segments_from_vf_onset.py --interval_start 120 --interval_end 180

Binary Classification with Transformer Model

This project demonstrates a binary classification example using a Transformer model. The model is implemented with TensorFlow and is designed for processing text or sequence data.

Key Components

1. Data Loading and Preprocessing

   • Load .npy files and separate data from labels.
   • Split the dataset into training, validation, and test sets, and save them.

2. Model Configuration

   • Transformer Model: Composed of Transformer blocks including Multi-Head Attention, Positionwise Feedforward Network, and Positional Encoding.
   • Training and Evaluation: Train and evaluate the model, and visualize the ROC curve.

3. Running in Jupyter Notebook

   • Includes DebugModel class for debugging intermediate outputs of the model.

Installation and Requirements

Before running the project, install the necessary libraries and packages:

pip install tensorflow numpy scikit-learn matplotlib pillow