[Paper Review] Topological Data Analysis for Arrhythmia Detection through Modular Neural Networks

Meryll Dindin, Yuhei Umeda+1 more|arXiv (Cornell University)|Jun 13, 2019

Topological and Geometric Data AnalysisComputer Science35 references46 citations

TL;DR

The paper proposes a modular multi-channel neural network for ECG-based arrhythmia detection and classification that incorporates topological data analysis (TDA) to improve generalization across unseen patients, achieving competitive results with state-of-the-art methods.

ABSTRACT

This paper presents an innovative and generic deep learning approach to monitor heart conditions from ECG signals.We focus our attention on both the detection and classification of abnormal heartbeats, known as arrhythmia. We strongly insist on generalization throughout the construction of a deep-learning model that turns out to be effective for new unseen patient. The novelty of our approach relies on the use of topological data analysis as basis of our multichannel architecture, to diminish the bias due to individual differences. We show that our structure reaches the performances of the state-of-the-art methods regarding arrhythmia detection and classification.

Motivation & Objective

• Motivate robust generalization in arrhythmia detection/classification across new unseen patients.
• Integrate topological data analysis with deep learning to capture geometry of ECG signals independent of individual differences.
• Develop a modular architecture combining TDA features, auto-encoders, and conventional DL channels.
• Evaluate on open-source Physionet ECG datasets with patient-based cross-validation.
• Provide benchmarks showing the value of TDA in improving generalization for multi-class arrhythmia classification.

Proposed method

• Introduce persistence barcodes and Betti curves to represent 1D ECG signals in a way robust to time rescaling and amplitude changes.
• Compute TDA features from ECG via persistent homology of sub-level and upper-level sets of the time series.
• Use Betti curves as stable, input-ready representations for 1D signals in a CNN-compatible pipeline.
• Incorporate an auto-encoder trained on normal beats to address data imbalance and provide reconstruction-based features.
• Build a modular deep-learning architecture with channels for TDA, auto-encoder outputs, Fourier features, and standard CNN/FC pathways.
• Apply annealed dropout to stabilize training and improve generalization.
• Evaluate using patient-based cross-validation to test generalization to unseen patients.

Experimental results

Research questions

• RQ1 Does adding a TDA-based channel improve arrhythmia detection and classification performance on unseen patients?
• RQ2 Does the combination of TDA features with auto-encoder-derived representations enhance generalization under data imbalance?
• RQ3 How does the proposed modular architecture compare to state-of-the-art methods on open ECG datasets in terms of generalization to new patients?
• RQ4 What is the impact of different channel configurations (with/without TDA) on binary detection and multi-class classification performance?

Key findings

• TDA improves both arrhythmia detection and classification, with notable gains in multi-class classification.
• The encoder and TDA channels provide the greatest impact, especially for 13-class classification.
• Mean validation accuracy for 13-class classification reached 97.3%, with test accuracy of 80.5% under patient-based cross-validation.
• Arrhythmia detection achieved mean validation accuracy around 98% and test accuracy around 90% under cross-validation.
• The approach achieved competitive results relative to benchmarks on PVC detection and multi-class tasks, highlighting improved generalization via TDA and auto-encoder fusion.