Overview

Assignment Specification

Overview

Electrocardiogram (ECG) signals aretime-series datathat visually represent the electrical impulses generated by the myocardium. Many physicians rely on ECG signals as their primary screening tool for the detection and diagnosis of cardiac diseases.

Figure 1depicts a graphical representation of the cardiac activity observed in heartbeats through an ECG. This recording is obtained by placing electrodes on a patient's chest during testing. As shown in the figure, an ECG signal comprises five distinct morphological waveforms: P, Q, R, S, and T waves. The PQRST waveform results from the depolarization (working phase) and repolarization (resting phase) activities occurring in each cardiac cycle. There are three main components in an ECG:

P wave, indicative of atrial depolarization, is recognized as the natural pacemaker responsible for pulse generation.

QRS complex, encompassing the Q wave, R wave, and S wave, represents ventricular depolarization generated by the atrioventricular node.

T wavesignifies ventricular repolarization.

A full cardiac cycle consists of the sequential depolarization and repolarization, progressing from the atrium to the ventricle. A normal ECG, exemplified inFigure 1, reveals a steady heart rhythm. Conversely, an abnormal ECG lacks this consistent rhythmic pattern, signaling potential issues such as damage to the heart muscle or irregularities in the heart's rhythm. Medical professionals identify these abnormalities in an ECG report by scrutinizing the rhythmic patterns of the ECG signal, drawing upon their expertise and experience.

The growth of artificial intelligence promises new opportunities in the development of automatic analysis and diagnostic techniques for ECG. These advancements not only alleviate the workload on doctors but also improve the efficiency and accuracy of ECG diagnosis substantially. Existing machine learning-based methods for anomaly detection are typically categorized into three groups:supervised,semi-supervised, andunsupervised learning. From a data science perspective, ECG data can be approached as time-series data. A summary and comprehensive evaluation of existing methods for anomaly detection in time seriescan be found athttps://dl.acm.org/doi/10.14778/3538598.3538602Project Task

This project is centered around the utilization of machine learning (ML)-based methods for ECG anomaly detection. The students are expected to:

determinean appropriate typeof ML-based methods through the analysis of the given ECG dataset,

employat least twoML techniques to detect anomalies in the ECG data, and

conductcomparative analysisanddraw insightsfrom the analysis.

Please note that the assessment of the project task will emphasize (1) the explanation of the rationale behind your choice of ML algorithms and the design of your methods and (2) the interpretation of analysis results. It goes beyond a judgement solely based on the performance of the algorithms or methods themselves. In other words, the key aspects under scrutiny during the assessment of the project task will behowandwhyyou craft the algorithmic approach, refine hyperparameters, andwhatvaluable insights you gain from the analysis.

Dataset

PhysioNet/CinC Challenge 2017is a real-life ECG recording set that was initially used to solve a classification problem - to classify whether a single short ECG recording shows anormal sinus rhythm, anatrial fibrillation, analternative rhythm, or it is toonoisyto be classified. The ECG recordings were sampled at 300 Hertz and all data is provided in MATLAB V4 WFDB-compliant format, each including a.matfile containing the ECG and a.heafile containing the waveform information.

In this project, we are interested in anomaly detection, where thenormal sinus rhythmis deemed as normal rhythmic patterns, and the other three classes, i.e.atrial fibrillation,other rhythm(any rhythm other thannormal sinus rhythmandatrial fibrillation) andnoise, are consideredabnormal. We updated theoriginal dataset and divided the entire dataset into:

Training set: 4060 recordings of normal sinus rhythm (labeledN)

Test set: 1016 recordings of normal sinus rhythm (labeledN), 758 recordings of atrial fibrillation (labeledAF), 2415 recordings of other rhythms (labeledO), and 279 recordings of noise (labeled~).

Where to start

Understand ML-based methods to handle ECG recordings.

Existing work from PhysioNet/CinC Challenge 2017 (https://doi.org/10.22489/CinC.2017.360-239Links to an external site.)

Existing work on detecting abnormal ECG recordings (https://doi.org/10.1016/j.engappai.2023.106484Links to an external site.)

Review prior work

Existing work on time-series anomaly detection. Two key literatures in the domain are listed here (https://dl.acm.org/doi/10.14778/3538598.3538602Links to an external site.,https://arxiv.org/abs/2211.05244Links to an external site.)

Tools and platforms

Any programming language and platform suitable for solving data science problems can be applied in the project.

To process ECG recordings, there are two approaches available:

Treat the.matfile as a standard MATLAB file and transform it into an array. For instance, employ 'scipy.io.loadmat()'function in Python package SciPy.

Utilize the WFDB package (https://wfdb.ioLinks to an external site.) to read ECG and employ in-package methods for converting the data into formats compatible with your models.

Note: ONLY using readily available functionalities offered in the existing anomaly detection benchmarking toolkits to conduct this project will UNLIKELY be sufficient to meet the expectations of the project. You need to replicate the existing methods by building your own pipeline.

Research questions:

1How effective is anomaly detection model trained solely on normal ECG rhythms in identifying abnormal rhythms.

2How do autoencoder, Variational autoencoder and RNN models perform in anomaly detection across different types of ECG anomalies?

What to do and deliverables

Task 1

This task will run three machine learning models to answer the research question

Autoencoder

Variational autoencoder

LSTM (a type of RNN that handles log-term dependencies)

Identify which models perform better generally and specifically in detecting specific types of anomalies and explore the reasons behind their performance.

NB: The data for this analysis contains all normal rhythms in the training dataset and contains both normal and abnormal in the testing dataset so it is not a classification or regression task rather an anomaly detection task.

Task 2

Write in details the methodology in carrying out this analysis. Steps for preprocessing data must be detailed and the reason for taking such steps must be detailed. Steps employed in running the three models must be detailed. In pre-processing data note this To process ECG recordings, there are two approaches available:

Treat the.matfile as a standard MATLAB file and transform it into an array. For instance, employ 'scipy.io.loadmat()'function in Python package SciPy.

Utilize the WFDB package (https://wfdb.ioLinks to an external site.) to read ECG and employ in-package methods for converting the data into formats compatible with your models.

Task 3

Result must be clearly presented and discussed, drawing insights from each analysis, also make a comparative analysis of the three models used. Compare the performance of each models based on various evaluation metrics

Evaluation metrics(accuracy, precision, recall and F1-score) must be clearly discussed. You can also write whether they agree with literature and provide your references accordingly.

Reconstruction error for autoencoder and VAE

Computational efficiency(training time, memory usage).

Task 4

Write a comprehensive conclusion of the work done and recommendation for future work.

Discuss the implications of the results for anomaly detection in ECG signals and potential applications in healthcare.

Highlight the limitations of the study and suggest avenues for future research to address them.

Task 5

Visualization and Interpretation

Visualize the constructed ECG signals from autoencoder and VAE to assess the quality of reconstructions.

Visualize the output of the RNN model to understand its predictions and identify patterns or anomalies in the ECG data.

Interpret the learned representations from each model to gain insights into the underlying structure of the data and the models decision -making processes.

Task 6

Prepare a PowerPoint presentation for the work you have done in a way that it will tell stories.