[Paper Review] Deep learning neural nets for detecting heart activity

J.M. Horvath, Lu Shien+4 more|arXiv (Cornell University)|Jan 16, 2019

ECG Monitoring and Analysis8 references4 citations

TL;DR

This study proposes a deep learning approach using feedforward and time-delayed neural networks to reconstruct heart surface potentials (HSPs) from body surface potential (BSP) measurements, addressing the ill-posed inverse problem of electrocardiography. By identifying optimal BSP electrode combinations via a meta-heuristic method, the model achieved high correlation (≥0.8) in reconstructing 80 out of 108 HSPs, demonstrating practical feasibility for non-invasive cardiac diagnosis.

ABSTRACT

The prediction of heart surface potentials using measurements from the body's surface is known as the inverse problem of electrocardiography. It is an ill-posed problem due to the multiple factors that affect the heart signal as it propagates through the body. This report details research performed into a machine learning solution to signal reconstruction as well as an analysis of optimal torso electrode positioning for prediction involving different areas of the heart. The dataset contains simultaneous measurements from a large number of body surface potential (BSP) and heart surface potential (HSP) electrodes, as well as their geometric locations, recorded from an experiment using a human model. Initially, Time Delayed Neural Nets were trained and tested across all BSP to HSP relationships resulting in a slight trend of increased reconstruction correlation with decreased separation of electrodes. However, the TDNNs had overfitted to the data and failed to predict alternate heartbeat pacings. Feed Forward Neural Nets (FFNNs) were tested in a many BSP to many HSP prediction method. Again overfitting occurred. To reduce overfitting, the number of training signals was reduced by analysing the optimal training BSPs for each HSP when using basic perceptrons. This analysis involved repeat sampling and ranking of different BSP combinations, initially, using a Monte Carlo approximation, until being replaced with a meta-heuristic which increased the yield of successful BSP combinations. Successful reconstructions across heartbeat pacings were produced using these optimal BSP combinations for 80 of the 108 HSPs, and future work exists for the testing of this method of prediction using real patient data.

Motivation & Objective

• To address the ill-posed inverse problem of electrocardiography by reconstructing heart surface potentials (HSPs) from body surface potential (BSP) measurements.
• To investigate the impact of electrode positioning on reconstruction accuracy, particularly the spatial locality between BSPs and HSPs.
• To reduce overfitting in neural network models by selecting optimal subsets of BSP electrodes for training.
• To develop a meta-heuristic method for identifying high-yield BSP combinations that improve generalization across different heartbeats.
• To validate the approach on a human model dataset and prepare for future testing on real patient data.

Proposed method

• Trained feedforward neural networks (FFNNs) and time-delayed neural networks (TDNNs) to map BSPs to HSPs, minimizing mean square error using the Levenberg algorithm.
• Applied sigmoid activation functions and used cross-validation to optimize hidden layer neuron counts and delay window sizes (0–20) in TDNNs.
• Implemented a Monte Carlo approximation followed by a meta-heuristic algorithm to rank and select optimal BSP combinations for each HSP.
• Used geometric projection mapping to translate 2D electrode positions on the torso to 3D anatomical coordinates using CT scan data and patient-specific torso dimensions.
• Reversed the projection process to recommend practical electrode placements on a patient’s torso for optimal HSP reconstruction.
• Normalised electrode data using geometric and signal-based methods to improve model robustness across anatomical variations.

Experimental results

Research questions

• RQ1 Does the spatial locality between body surface electrodes and heart surface potentials improve reconstruction accuracy in the inverse problem of electrocardiography?
• RQ2 Can a meta-heuristic approach effectively identify optimal subsets of body surface electrodes that reduce overfitting and improve generalization across different heartbeats?
• RQ3 To what extent can deep neural networks reconstruct HSP signals from BSP data with high correlation, particularly when trained on reduced, optimised electrode sets?
• RQ4 How do different electrode configurations (e.g., anterior, posterior, lateral) compare in their ability to reconstruct specific HSPs?
• RQ5 Can the identified optimal electrode combinations be practically applied in a clinical setting using real patient data?

Key findings

• The spatial locality between BSPs and HSPs significantly improves reconstruction correlation, with closer electrodes yielding better results.
• Using a meta-heuristic method to select optimal BSP combinations reduced overfitting and enabled successful reconstruction of HSP signals across different heartbeats.
• For 80 out of 108 HSPs, the model achieved a reconstruction correlation of at least 0.8 using optimised BSP combinations.
• The optimal electrode combinations ranged from 2 to 24 electrodes, which is clinically feasible compared to standard 12-lead ECGs.
• The method successfully reconstructed signals even when the heart pacing varied, indicating robustness to beat-to-beat changes.
• The geometric projection and normalisation techniques enabled practical translation of electrode recommendations to patient-specific torso configurations.