Cardiac arrhythmia detection using artificial intelligence

Cardiac arrhythmia detection using artificial intelligence

1. Paper Title : Cross-Domain Cardiac Arrhythmia detection using Deep Transfer Learning on ECG Signals Name: KM Namrata Bose Kshitij Goyal Soumyadeep Manna Dept. of Computer Science, Amity School Of Engineering and Technology, Amity University, Uttar Pradesh

2. Challenges in Detecting Cardiac Arrhythmias Cardiac arrhythmias are abnormal heart rhythms that can disrupt blood flow. Accurate detection is crucial for timely diagnosis and treatment. Challenges in ECG signal analysis: Noise (power line interference) Artifacts (muscle movement) Baseline wander Subtle arrhythmias with minor ECG variations can be missed visually. Irregular ECG Waveform Doctor looking at the ECG of a Patient Heart with Regular ECG Signal

3. Understanding the ECG Waveform The ECG waveform is a graphical representation of the electrical activity of the heart during a heartbeat. It consists of distinct waves corresponding to different stages of the cardiac cycle: P wave: Atrial depolarization (contraction) QRS complex: Ventricular depolarization T wave: Ventricular repolarization Deep learning models can extract features from ECG waveforms for arrhythmia classification. ECG Waveform

4. Dataset Description MIT-BIH arrhythmia database: This dataset is a collection of ECG recordings from 47 patients used to train and test machine learning models for arrhythmia detection. Data characteristics: Source: Beth Israel Hospital, Boston Annotations: Over 109,000 annotations for different cardiac arrhythmia types by cardiologists. Availability: Freely accessible through Physionet ATM Bank. Data formatting: The data is formatted and processed according to AAMI EC57 standards. Class imbalance: The dataset has a significant imbalance, with "Normal" being the most frequent class and "Fusion Beat" being the least frequent. Images showing ECG Signals Table showing Dataset Classification

5. Data Pre-Processing Total data points: 109,446 annotations for cardiac arrhythmia. Split: 80% (87,554) for training and 20% (21,892) for testing. Data format: Each row represents an annotation with 187 columns. Last column: Class label (N, S, V, F, or Q) Remaining 186 columns: Normalized ECG data (values between 0 and 1) with zero padding for missing values. Class imbalance handling: The original data has an imbalance between classes (some classes have significantly fewer examples). Solution: Oversampling is used to address this issue. Oversampling increases the number of data points for minority classes in the training data. This helps the model learn better from these classes and potentially improve classification performance for them. Illustration of Data Pre-Processing

6. Proposed Network Architecture Proposed Network Architecture

7. Proposed Network Architecture (continue…) We propose a deep transfer learning framework for automatic cardiac arrhythmia detection using ECG signals. The architecture leverages a pre-trained 1D convolutional neural network (CNN) model. The CNN extracts features from the ECG signal for arrhythmia classification. The architecture consists of three main stages: Convolutional layers: Extract features from the ECG signal (e.g., patterns, trends). Pooling layers: Reduce dimensionality and capture important features. Fully connected layers: Combine features and classify the ECG signal into arrhythmia classes. Additional components (optional): Depending on your architecture, you might have elements like: Batch normalization: Improves training stability. Dropout layers: Prevents overfitting. Activation functions: Introduce non-linearity for complex decision making.

8. Result Resampling techniques address data imbalance and improve model generalizability. This experiment compares the impact of resampling (e.g., oversampling or undersampling ) on the test set accuracy. The graph illustrates the accuracy achieved on the test set with and without resampling. The model trained with resampling achieves a significant improvement in test set accuracy, reaching 97.74% compared to 92.7% without resampling. This highlights the benefit of resampling for imbalanced datasets. Impact of Resampling on Accuracy Comparison of accuracy with & without Resampling

9. Result Performance of the Pre-trained Model This slide explores the training performance of the pre-trained CNN model used for transfer learning. The graph depicts the loss and accuracy curves across training epochs. Loss curve (typically decreasing): Indicates the model's ability to learn and minimize the loss function during training. Accuracy curve (typically increasing): Represents the model's capability to correctly classify ECG signals. We analyze these curves to understand the pre-trained model's suitability for transfer learning. Accuracy Curve Loss Curve

10. Result Comparison of Accuracy with and Without Transfer Learning This slide compares the test set accuracy achieved with and without transfer learning. The model trained using transfer learning significantly outperforms the model trained from scratch. The graph (or table) illustrates the accuracy values for both scenarios. In this experiment, transfer learning led to an improvement in test set accuracy from 85.23% (without transfer learning) to 97.3% (with transfer learning) . This demonstrates the effectiveness of transfer leaarning for improving arrhythmia classification accuracy. Comparison of accuracy with & without Resampling

11. Result Performance Comparison of Different Optimizers and Dropout Rates Performance Comparison of Different Optimizers and Dropout Rates This slide explores the impact of hyperparameter tuning on the transfer learning model's performance. Hyperparameters control the learning process of the model (e.g., dropout rate, optimizer). We evaluated the model's performance using various combinations of dropout rate and optimizer. The table depicts the test set accuracy achieved for each combination. As you can see in the table, the best performing configuration achieves a test set accuracy of 97.74% . This configuration likely utilizes the Adam optimizer with a dropout rate of 30% (based on your previous description).

12. Conclusion Key Findings and Future Directions Transfer Learning Success: Pre-trained networks with data resampling achieved high accuracy, suggesting promise for reliable arrhythmia detection systems. Validation Needs: Future work should include validation on a larger, more diverse dataset for generalizability. Real-time implementation with wearable devices is another crucial step. Improved Patient Care Potential: By exploring these directions, pre-trained networks can become even more valuable for arrhythmia detection and patient treatment. Major Contributions Novel Framework: This research proposes a novel framework for automatic cardiac arrhythmia detection using deep transfer learning with ECG signals. Addressing Imbalance: The study tackles class imbalance by employing up-sampling techniques. Transfer Layer Analysis: Performance analysis with different transferring layers optimizes the transfer learning approach for this task. ECG Signal with heart

13. Thank You