Neural Networks for Heart Disease Detection via ECG

Leveraging Neural Networks for Early Detection of Cardiac Disease: An ECG Analysis Approach

Project Overview

The goal of the project is to automatically identify heart conditions by analyzing electrocardiogram (ECG) data. It makes use of a dataset made up of 827 recordings, each of which has 4096 samples, or cycles, of 12 leads in the ECG. The main objective is to create a predictive model that can use ECG data analysis to identify symptoms of heart illnesses.

 

Electrocardiography is the technique of using electrodes applied to the skin to create an electrocardiogram, often known as an EKG or ECG, which is a graphical representation of the voltage versus time of the heart's electrical activity. 

How is the ECG recorded?

The electrical depolarization of the heart starts at the atria and progresses to the ventricles via the interventricular septum (IVS). Because of the physical location of the heart and the bigger muscle mass of the left ventricle, this depolarization usually occurs from the top (superior) to the bottom (inferior), with a leftward orientation. The electrical axis is the name given to this directed flow of electrical activity.

A fundamental idea in ECG recording is that an upward or positive deflection occurs when the depolarization wave approaches a recording lead. However, it produces a negative or downward deflection as it moves away from a recording lead. This foundational knowledge directs the interpretation of ECG signals, enabling the detection of irregularities and anomalies in the heart.

Project Goals and Objectives

The objective of this study is to create a system that can evaluate electrocardiogram (ECG) data and identify cardiac problems in patients by using neural network technology. By incorporating generative AI techniques, the system aims to enhance data interpretation and improve the accuracy of early detection, allowing for prompt intervention and better patient outcomes.

• Focus on Early Detection: The project's objective is to find possible cardiac disorders as soon as possible. Early identification enables prompt intervention and therapy, which may enhance patient outcomes and lower the chance of heart disease complications.
• ECG Data Analysis: The project's main information source is ECG recordings. Heart electrical activity is recorded by ECGs, and anomalies in these signals may point to underlying cardiac issues.

Project Challenges

There were several challenges that the project faced, including:

• Efficiently managing and analyzing massive amounts of ECG data.  
• Identifying relevant characteristics in ECG data that point to cardiac anomalies.  
• Creating a neural network architecture that can analyze ECG data and identify intricate patterns and relationships.  
• Ensuring the model's universality and robustness across various patient groups and data fluctuations. 

Technologies Used

Several technologies were used in the project, such as:  

• Python: For developing, evaluating, and preparing data models.  
• Neural network model construction and training is done using deep learning frameworks like TensorFlow or PyTorch.  
• HDF5: A file format used to store and arrange ECG data uniformly.
• Numerous libraries for signal processing and machine learning are used for data analysis and feature extraction. 

Predictive Model Development: Developing a machine learning model is the main goal of the project. The supplied ECG data and accompanying diagnoses of cardiac illness (present or missing) are used to train this model. The model gains the ability to recognize relationships and patterns in the ECG data connected to various cardiac diseases throughout training. 

Solutions we Provided

The following solutions were put into practice to deal with the challenges and meet the project's goals:

• A thorough examination of the ECG reveals pertinent characteristics linked to heart conditions.  
• Creation of a neural network architecture with accurate prediction-making abilities for processing ECG data.  
• Application of data augmentation strategies to improve the robustness and generalization of the model.  
• Performance metrics can be improved by model training and hyperparameter optimization. 

Breakdown of the Dataset

• Recordings: 827 distinct ECG recordings make up the dataset that the system examines.  
• Samples per Recording: 4096 samples, or individual cardiac electrical activity cycles, make up each recording.  
• 12 Leads: Data is collected from 12 distinct electrode sites on the limbs and chest during each recording. These distinct leads offer diverse viewpoints on the electrical activity of the heart, providing a more complete picture. 

We discovered after carefully examining the data that any waves that don't exhibit regularity have a higher likelihood of obstruction and anomalies.

The ECG is used to extract the following features: rhythm, rate, axis, PR interval, Q wave, QRS complex, QT interval, ST interval, T wave, hour mean, hour standard deviation, hour v mean, and hour v standard deviation.  

Heart Rate, or HR  

Heart Rate Variability (HRV): P, Q, R, S, and T correspond to the ECG 

Understanding the DataSet 

Heart's Electrical Fingerprint: ECG Leads

The ECG leads function like several cameras, recording the electrical activity of the heart from various perspectives. Through examination of these "views" (standard, enhanced limb, and chest leads), medical professionals might see trends that may indicate cardiac problems: 

• All leads exhibit sinus bradycardia, or a slow heartbeat.  
• All leads exhibit sinus tachycardia, or an accelerated heartbeat.  
• In all leads, delayed conduction (1st Degree AV Block) manifests as an extended PR interval.  
• Blockages (RBBB & LBBB) exhibit distinct wave patterns based on their specific position. 
• Atrial fibrillation, or irregular heartbeat, throws off the regular rhythm in all leads. 
  • All these observations were discovered by reading various research papers. We therefore deduced that variations in the ECG are associated with an increased risk of blockage or other cardiac problems.  
  • All these features were taken out of the ECG and used to train our model.

An explanation of model architecture  

The "model.py" file contains the definition of the analysis model. The model's functions are broken down as follows:

• Input: A portion of the ECG data that includes:  
• The quantity of ECG recordings processed in a batch (N) is known as batch size.  
• The number of electrical measurements per recording (sampled at 400Hz) is shown by the data points (4096).  
• Leads (12): The twelve distinct measures of the ECG leads.  
• Data Format: 32-bit floating-point values scaled to 1e-4 volts are used to represent each ECG data point. Before feeding your data into the model, multiply it by 1000 if it is in volts.  
• Output: For each of the six potential cardiac anomalies, the model generates a likelihood score.

• The train.py method is used to train the neural network 
• The script takes the path to the HDF5 file containing the ECG tracings and the path to the 
• CSV file with the corresponding labels. 
• Pre-trained models obtained using this script are available for download. 
• The predict.py script is used for generating predictions on a given data set. 
• The script takes the path to the HDF5 file with ECG tracings, the path to the trained 
• model, and optionally, the path to the output file for saving the predictions. 
• The generate figures and tables py.script is used to create images. 
• It analyzes predictions generated from the models and generates visualizations and performance matrix 

Result 

 Test Accuracy: 84.3375% 

Confusion Matrix

 

Machine Learning Model

This is a computer program that has been trained on a large dataset of ECG readings that have previously been annotated with the patient's diagnosis of heart illness. The program gains the ability to recognize patterns in ECG data that are frequently connected to cardiac issues by analyzing these examples.

Neural Network 
This machine learning model is modeled after the human brain. It's like having an ultra-skilled investigator who can simultaneously examine the ECG data from multiple perspectives, uncovering even the most minute hints. 

Confusion Matrix 
The model indicates whether an illness is likely or unlikely after analyzing a fresh ECG report. Comparable to a report card, the confusion matrix displays the model's performance. It shows how often the model was correct or incorrect by comparing its estimations to the actual diagnoses. This enables us to assess the model's performance in identifying cardiac illness and identify areas for potential improvement.

Benefits Gained by this project

The effort resulted in several advantages, such as: 

•  Improved heart illness early detection, which enables prompt intervention and treatment.  
• Increased effectiveness in evaluating substantial amounts of ECG data, lightening the workload for medical practitioners.  
• Improved disease prediction accuracy with fewer false positives and false negatives.  
• Customized therapy by offering recommendations that are specifically designed based on each patient's ECG profile.  
• Better delivery of healthcare as this technology may make cardiac screening more accessible, especially in places with little access to medical resources.  
• Reduced healthcare costs as the costs of standard diagnostic tests may be lowered by automating disease diagnosis. 

Conclusion

The effective creation of an automated neural network system for identifying heart disorders shows how AI-powered image recognition with Convolutional Neural Networks (CNNs) can transform the field of medical diagnosis. Using data-driven strategies and cutting-edge technology, medical procedures can be streamlined, and patient outcomes can be improved.

Future Project Scalability

• Automated analysis of ECG data during routine checks may be possible with integration with electronic health record (EHR) systems.  
• The model could be improved to identify the precise kind of heart ailment in addition to identifying its presence.  
• Patients could be able to record and upload ECG data via mobile applications so that the system can analyze it remotely.