mohit-bags/ECG-523

Cardiologist-level arrhythmia detection and classification using deep neural networks.

ECG Arrhythmia Classification with Deep Neural Networks

Team Members

• Anthony Sayegh (asayegh@bu.edu)
• Alvaro Carrascosa Penabad (varo@bu.edu)

Included Files

• Models --> Stores the models and their corresponding weights
• Data --> Contains the MIT-BIH data and the links to the additional required data (Too big for GitHub)
• Utilities --> Contains helper functions provided for the Physionet challenge
• Slides --> Contains slides for the in-class presentation
• Images --> Contains data visualizations and other generated useful representations and results

Goals

Explore the following types of models for Arrythmia Classification using Deep Neural Networks

• Detector Scale Algorithm(CNN) : Simple architecture, basic detection of irregularities.
• Diagnosis Scale Algorithm(ResNet) : More classes, but utilizes a more complex architecture in addition to having more parameters.

Example of Input file (12-lead ECG)

Traditional 12-lead ECG Plot Alternate 12-lead ECG Plot(sample# vs. ADC count)

Datasets

• MIT-BIH Arrhythmia dataset <-- Used for Detector Model

• Physionet Challenge 2020 (SNOMED Mapping) <-- Used for Diagnosis Model

  • China 12-Lead ECG Challenge Database
  • China Physiological Signal Challenge in 2018
  • Georgia 12-Lead ECG Challenge Database
  • PTB Diagnosis ECG Database
  • PTB-XL electrocardiography Database
  • St Petersburg INCART 12-lead Arrhythmia Dataset

  Total of 43101 Recordings (See distribution of classes in data below)

Results

• CNN Model (Detection)
  • F1 Score --> 0.87 (10 epochs)
  • Number of Parameters --> ~ 15,000
  • Confusion Matrix

List of Classes in Confusion Matrix

• N = Normal Beat
• S = Supraventricular
• V = Ventricular
• F = Fusion Beats
• Q = Unknown Beats

---

• ResNet Model (Diagnosis)
  • Number of Parameters --> ~ 500,000
  • Confusion Matrix

List of Classes in Confusion Matrix

• IAVB = 1st Degree AV Block
• AF = Atrial Fibrillation
• AFL = Atrial Flutter
• Brady = Bradycardia
• CRBBB = Complete Right Bundle Branch Block
• IRBBB = Incomplete Right Bundle Branch Block
• LAnFB = Left Anterior Fascicular Block
• LAD = Left Axis Deviation
• LBBB = Left Bundle Branch Block
• LQRSV = Low QRS Voltage
• NSIVCB = Nonspecific Intraventricular Conduction Disorder
• PR = Pacing Rhythm
• PAC = Premature Atrial Contraction
• PVC = Premature Ventricular Contraction
• PR = Prolonged PR Interval
• LQT = Prolonged QT Interval
• QAb = Qwave Abnormal
• RAD = Right Axis Deviation
• RBBB = Right Bundle Branch Block
• SA = Sinus Arrhythmia
• SB = Sinus Bradycardia
• SNR = Sinus Rhythm
• STach = Sinus Tachycardia
• SVPB = Supraventricular Premature Beats
• TAb = T Wave Abnormal
• TInv = T Wave Inversion
• VPB = Ventricular Premature Beats

---

Link to Slides

• https://github.com/varocarras/ECG-523/blob/main/Slides/CS523%20Project.pdf

Demo Code

• DEMO link
• This demo is located in the SCC at sftp://varo:@scc1.bu.edu/ECG-523/ , all the data has been downloaded to the SCC.
• The demo code allows for training the model as well as loading its weights to predict a rhythm for any given waveform.
• Run the Demo.ipynb to generate visualizations and show an example of predicting a rhythm for a particular patient.
• In order to retrain the model:
  • Run "python train.py" (root folder file) or uncomment the training code from the Demo.ipynb
  • Weights will save to /Models/resnet_model.h5 and will load from there to test
  • We left the weights there in case you dont want to train the model again

References

1. B. -J. Singstad and C. Tronstad, "Convolutional Neural Network and Rule-Based Algorithms for Classifying 12-lead ECGs," 2020 Computing in Cardiology, 2020, pp. 1-4, doi: 10.22489/CinC.2020.227. --> (https://ieeexplore.ieee.org/document/9344421)

2. Cardiologist Level Aryhthmia Detection with Convolutional Neural Networks --> (https://arxiv.org/abs/1707.01836)

3. MIT-BIH Data in CSV format --> (https://www.kaggle.com/shayanfazeli/heartbeat)

4. Detector Model Adaptation --> (https://www.kaggle.com/gregoiredc/arrhythmia-on-ecg-classification-using-cnn)