Source-linked AI summary
Designing ECG Monitoring Healthcare System with Federated Transfer Learning and Explainable AI
Ali Raza, Kim Phuc Tran, Ludovic Koehl, Shujun Li
TL;DR
Deep-learning ECG classification is constrained by limited siloed data, privacy concerns, noisy real-time signals, and limited explainability. The paper proposes a federated, transfer-learning, CNN-based framework with denoising and XAI components; it reports improved federated classification performance while adding explainability and privacy features.
Problem
ECG deep-learning systems face data-availability and privacy challenges in centralized settings, while noisy real-time data and limited explainability constrain healthcare use.
Method
The framework combines federated learning, transfer learning, a one-dimensional CNN denoising autoencoder, a CNN classifier, and an XAI module for ECG classification and explanation.
Results
The proposed method outperformed other methods in the reported comparison while handling five classes in a federated architecture and providing explainability and data privacy.
Takeaways & Limitations
The framework provides an end-to-end approach for privacy-preserving, efficient, and interpretable ECG signal classification in federated settings.
Takeaways & Limitations
The distributed setting increases exposure to data-poisoning attacks, and the framework assumes homogeneous data and devices across distributed edges.
Abstract
from arXiv · showhide
Deep learning play a vital role in classifying different arrhythmias using the electrocardiography (ECG) data. Nevertheless, training deep learning models normally requires a large amount of data and it can lead to privacy concerns. Unfortunately, a large amount of healthcare data cannot be easily collected from a single silo. Additionally, deep learning models are like black-box, with no explainability of the predicted results, which is often required in clinical healthcare. This limits the application of deep learning in real-world health systems. In this paper, we design a new explainable artificial intelligence (XAI) based deep learning framework in a federated setting for ECG-based healthcare applications. The federated setting is used to solve issues such as data availability and privacy concerns. Furthermore, the proposed framework setting effectively classifies arrhythmia's using an autoencoder and a classifier, both based on a convolutional neural network (CNN). Additionally, we propose an XAI-based module on top of the proposed classifier to explain the classification results, which help clinical practitioners make quick and reliable decisions. The proposed framework was trained and tested using the MIT-BIH Arrhythmia database. The classifier achieved accuracy up to 94% and 98% for arrhythmia detection using noisy and clean data, respectively, with five-fold cross-validation.
1. Introduction
Healthcare deep learning faces data scarcity, privacy, and explainability challenges, especially for real-time ECG classification. The paper proposes a federated, explainable framework combining denoising, classification, and interpretation.
- Healthcare data are distributed across silos, while collecting them centrally raises privacy concerns and may leave individual sources insufficient for training.
- Deep learning predictions can be difficult for clinical practitioners to understand, limiting healthcare use when explanations are required.
- ECG classification is important, but existing approaches commonly rely on centralized, cleaner data and often lack explainability for noisy real-time signals.
- The proposed framework combines an autoencoder, classifier, and XAI module in a federated setting for ECG arrhythmia classification.
- The framework aggregates information from edge devices, supports relatively personalized learning through knowledge transfer, and provides interpretable predictions.
2. Related Work and Background
Related work establishes federated learning, autoencoders, transfer learning, and XAI as relevant tools for privacy-preserving and interpretable healthcare systems. The paper applies these ideas to distributed ECG analysis and denoising.
- Centralized healthcare learning can have limited real-world applicability because collecting data centrally raises privacy concerns.
- An autoencoder learns an encoded representation and reconstructs the input, reducing dimensionality while retaining salient information.
- The study uses a denoising autoencoder that adds stochastic noise and learns to recover the original signal, improving robustness and reducing identity mapping.
- Federated learning trains models across distributed devices while keeping training data local and aggregating model updates.
- Explainable artificial intelligence helps users understand how machine-learning systems produce decisions and supports transparency, trust, and outcome assessment.
3. The Proposed Framework
The proposed framework combines federated transfer learning with a denoising autoencoder, CNN classifier, and Grad-CAM-based XAI module for privacy-aware, interpretable ECG classification. It also reduces communication by sharing selected lower-layer weights while retaining higher-layer personalization.
- The framework comprises an autoencoder, classifier, XAI module, and federated learning process for personal healthcare.
- CNN-based Autoencoder: The denoising autoencoder uses six convolutional, three maxpooling, and three upsampling hidden layers divided into encoder and decoder parts.
- CNN-based Classifier: The CNN classifier uses four convolutional layers, three max pooling layers, two fully connected layers, and one softmax layer to classify five ECG classes.
- Transfer Learning: The classifier transfers the autoencoder encoder and freezes its first three convolutional layers to retain learned denoising parameters during classification.
- XAI with Grad-CAM: Grad-CAM computes class-specific gradients, weights feature maps, applies ReLU, and upsamples the resulting heatmap to localize ECG regions relevant to predictions.
- XAI with Grad-CAM: The XAI visualization overlaps the denoised ECG signal with the heatmap so practitioners can identify regions receiving prediction focus.
- Communication Cost Reduction and Privacy Enhancement: Layer selection shares only trainable lower-layer weights for aggregation, reducing communication while higher-layer weights remain localized for classification.
- Communication Cost Reduction and Privacy Enhancement: Sharing lower-layer weights is intended to enhance privacy because these layers encode more common, low-level features than higher layers.
4. Experimental Results
Experiments used the MIT-BIH Arrhythmia Database and evaluated both original and artificially noisy ECG data on local edge devices in a federated implementation. The reported setup also measured global-round training and prediction-XAI execution times.
- Dataset: The MIT-BIH Arrhythmia Database contains 48 half-hour two-channel ECG excerpts from 47 subjects and 109,446 samples.
- Dataset: 10-30% noise was introduced into the original dataset to simulate more realistic ECG data alongside the clean version.
- Implementation: The autoencoder and classifier were trained locally on three Raspberry Pi edge devices, with a workstation serving as the global server.
- Implementation: The classifier used batch size 100, learning rate 0.001, and 150 training epochs.
- Execution Time: 745 seconds were required on average to complete one global training round, while prediction and XAI results took 2.32 seconds on average.
5. Performance Analysis of the Proposed Method
The framework is evaluated through autoencoder reconstruction and arrhythmia classification using standard metrics on noisy and clean ECG data.
- Reconstruction of Autoencoder: The aggregated autoencoder achieved reconstruction MAE nearly equal to 0, indicating effective reconstruction of the original signal.Its reconstruction MAE was less than or nearly equal to that of each local autoencoder.
- Classification Performance: Classification performance was measured using accuracy, precision, recall, and F1-score.Accuracy measures overall performance, while the other metrics assess class-specific discrimination.
- Classification Performance: For multiclass classification, precision, recall, and F1-score are calculated using one-vs-rest binary classifiers for each class.Accuracy is defined using total correct predictions divided by total samples.
5.3. Qualitative Analysis
The qualitative analysis uses XAI visualizations to show which ECG heartbeat regions influence classification decisions and to assess their clinical interpretability.
- Qualitative Analysis: The visualizations show the importance assigned to each beat for selected ECG signal inputs.These results are intended to help clinical practitioners assess underlying health issues.
- Qualitative Analysis: The XAI module highlights clinically important ECG regions, including the p-wave, QRS complex, and T-wave.Red segments indicate heartbeat regions with greater influence on the classifier’s prediction.
- Qualitative Analysis: Heat maps should be cross-checked by clinicians with expert knowledge before being used in medical consultation.The paper explicitly advises against using the results for medical consultation without prior discussion with a clinical professional.
5.4. Comparison With Other State-of-the-Art
The proposed framework is compared with state-of-the-art ECG methods on classification performance and supported system properties.
- Comparison With Other State-of-the-Art: Table 2 reports classification performance for the proposed framework on the noisy dataset.The table includes binary-classifier metrics at three edge devices and the global server, plus five-class accuracy.
- Comparison With Other State-of-the-Art: The proposed method outperformed the compared methods while classifying five classes rather than the two or three classes used by some alternatives.The comparison uses noisy data containing 10%-30% injected noise.
- Comparison With Other State-of-the-Art: The framework combines federated architecture, explainability, privacy protection, raw-signal denoising, and classification.The compared methods do not provide all of these properties together according to the reported comparison.
5.5. Privacy Enhancement
The privacy enhancement strategy limits federated sharing to lower-level parameters, reducing exposure of data-specific information while also lowering communication costs.
- Privacy Enhancement: Higher-level classifier parameters may retain data-specific information that could allow recovery of some information from shared parameters.This is identified as a residual privacy concern in federated learning.
- Privacy Enhancement: The proposed framework shares only lower-level learned parameters between edge or local servers and the global server.These parameters are described as carrying more common, low-level, and less privacy-sensitive features.
- Privacy Enhancement: Sharing fewer parameters can enhance privacy and reduce communication costs simultaneously.The reduction follows from transmitting fewer parameters between local and global servers.
5.6. Communication Cost Reduction
This section defines the communication-cost measure for federated training and reports an 8.2% reduction from the proposed communication-cost method.
- TPC denotes the total number of parameters communicated between an edge device and the global server for one global round.
- The section presents equations for calculating TPC in the framework and with the proposed communication-cost reduction method.
- 8.2% is the reported communication-cost reduction achieved by the proposed method.
5.7. Time Complexity of proposed Algorithm
The proposed CNN-based algorithm has linear time complexity because convolution and pooling scale linearly with input features, while edge-node parallelism preserves this overall runtime.
- O(n) is the overall runtime of the CNN-based network when the edge node computes in parallel.
- O(n) time is required to convolve a fixed-size filter across an input with n features.
- Max-pooling and average-pooling operations take no more than a linear amount of time in the input size.
5.8. Limitations and Future work
The framework’s limitations concern security threats, assumptions of homogeneous data and devices, internal attacks, and dependence on the global server.
- Federated deployment increases exposure to data-poisoning attacks, requiring stronger data-integrity and authentication methods.
- Assuming homogeneous distributed data and devices may limit local-model generalizability and reduce aggregated-model accuracy.
- The framework does not consider malicious internal attackers inserting hidden backdoors into the joint global model.
- Future work: Future work aims to address these limitations to produce a more secure and robust framework.
6. Conclusions
The paper concludes with a privacy-preserving, efficient, and explainable federated framework for ECG classification, combining denoising, classification, and explanation. It reports improved classification performance and identifies broader healthcare applications as future directions.
- The framework combines a CNN-based federated autoencoder for raw ECG denoising with a classifier and an explainable module.
- The federated autoencoder reconstructed raw ECG signals effectively and improved overall performance in federated settings.
- The federated classifier improved edge-device classification performance and outperformed existing centralized and federated algorithms on the baseline database.
- The explainable module visually indicated that clinically meaningful ECG heartbeat segments were behind the classification results.
- The framework is intended to support participation by healthcare data owners, diagnostic assistance where cardiology access is scarce, and identification of arrhythmia-triggering patterns.
- Future work: Future work targets human activity recognition, anomaly detection in home care, other arrhythmias, and the framework’s stated limitations.