Advance Prediction of Ventricular Tachyarrhythmias using Patient Metadata and Multi-Task Networks

Marek Rei, Joshua Oppenheimer, Marek SirendiarXiv 2018 · OpenReview

Abstract

We describe a novel neural network architecture for the prediction of ventricular tachyarrhythmias (VTA). The model receives input features that capture the change in RR intervals and ectopic beats, along with features based on heart rate variability and frequency analysis. Patient age is also included as a trainable embedding, while the whole network is optimized with multi-task objectives. Each of these modifications provides a consistent improvement to the model performance, achieving 74.02% prediction accuracy and 77.22% specificity 60 seconds in advance of the episode.

Introduction

Ventricular tachyarrhythmias (VTA), including ventricular tachycardia and ventricular fibrillation, are primary drivers of in-hospital cardiac arrest. While current monitoring systems are effective at detecting these events as they occur, the ability to predict them in advance remains a significant clinical challenge. Predicting VTA even 60 seconds before onset provides a critical window for medical staff to prepare for resuscitation or administer preventative treatments. This research aims to extend the prediction horizon using standard ECG data and patient metadata.

Methods

The proposed model is a multi-layer neural network designed to process heart rate variability (HRV) and patient-specific metadata.

Feature Engineering and Windowing

Beyond standard HRV metrics like mean RR intervals and frequency-domain power, the model incorporates "windowed features." These features capture the dynamics of heart rhythm changes by calculating statistics across two consecutive windows. This approach allows the model to detect trends and accelerations in cardiac instability rather than relying on static snapshots.

RR interval plot showing the windowed regions used to calculate trend features.

Architecture and Multi-Task Learning

The network consists of three hidden layers with tanh activations. Two key innovations distinguish this architecture:

  1. Age Embeddings: Patient age is treated as a trainable embedding, allowing the model to learn non-linear interactions between age and cardiac risk profiles.
  2. Multi-Task Optimization: During training, the network is optimized to predict VTA alongside auxiliary tasks: Body Mass Index (BMI) and New York Heart Association Class (NYHAC). This multi-task approach forces the shared hidden layers to learn more robust representations of general cardiac health.
Neural network architecture featuring age embeddings and multi-task outputs.

Dataset

The model was trained and validated using the PhysioNet Malignant Ventricular Tachyarrhythmia Database (MVTDB), comprising records from 78 patients with ICDs. The dataset includes 135 pre-event records (VT/VF) and 126 control records, with each sequence containing 1024 RR intervals leading up to the event.

Results

The model's performance was evaluated using 10-fold cross-validation. The incremental addition of windowed features, age embeddings, and multi-task optimization led to steady improvements in predictive power.

Configuration Accuracy Specificity AUC
Baseline (Standard features) 62.26% 61.27% 68.55
+ Windowed Features 71.03% 74.21% 78.52
+ Age Embedding 73.52% 77.22% 79.65
+ Multi-task Optimization 74.02% 77.22% 78.77

The final configuration achieved a prediction accuracy of 74.02% and a specificity of 77.22% at a 60-second lead time. Notably, the introduction of windowed features provided the most significant performance boost, highlighting the importance of temporal trends over instantaneous metrics.

Discussion

This work demonstrates that combining temporal signal trends with patient metadata and multi-task learning significantly improves VTA prediction horizons. While the results are promising, the study is limited by a relatively small dataset of 78 patients. Future research should focus on validating these findings in larger, more diverse clinical populations and exploring the optimal balance between lead time and predictive accuracy for real-world hospital deployment.

Summary updated

Abstract

We describe a novel neural network architecture for the prediction of ventricular tachyarrhythmias. The model receives input features that capture the change in RR intervals and ectopic beats, along with features based on heart rate variability and frequency analysis. Patient age is also included as a trainable embedding, while the whole network is optimized with multi-task objectives. Each of these modifications provides a consistent improvement to the model performance, achieving 74.02% prediction accuracy and 77.22% specificity 60 seconds in advance of the episode.
BibTeX
@misc{rei2018advance,
  author = {Marek Rei and Joshua Oppenheimer and Marek Sirendi},
  title = {{Advance Prediction of Ventricular Tachyarrhythmias using Patient Metadata and Multi-Task Networks}},
  year = {2018},
  eprint = {1811.12938},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.12938},
}
Figure