The Lock Generative Adversarial Network for Medical Waveform Anomaly Detection

TL;DR

This paper introduces the Lock GAN, a novel architecture for detecting anomalies in medical waveform data, demonstrating improved or comparable performance across ventilator and ECG datasets.

Contribution

The paper presents a new Lock GAN architecture with alternating optimization and synthetic minority oversampling for enhanced anomaly detection in medical waveforms.

Findings

Outperforms or matches state-of-the-art on ventilator asynchrony data

Achieves superior detection in ECG datasets

Demonstrates robustness across multiple medical waveform types

Abstract

Waveform signal analysis is a complex and important task in medical care. For example, mechanical ventilators are critical life-support machines, but they can cause serious injury to patients if they are out of synchronization with the patients' own breathing reflex. This asynchrony is revealed by the waveforms showing flow and pressure histories. Likewise, electrocardiograms record the electrical activity of a patients' heart as a set of waveforms, and anomalous waveforms can reveal important disease states. In both cases, subtle variations in a complex waveform are important information for patient care; signals which may be missed or mis-interpreted by human caregivers. We report on the design of a novel Lock Generative Adversarial Network architecture for anomaly detection in raw or summarized medical waveform data. The proposed architecture uses alternating optimization of the generator and discriminator networks to solve the convergence dilemma. Furthermore, the fidelity of the generator networks' outputs to the actual distribution of anomalous data is improved via synthetic minority oversampling. We evaluate this new architecture on one ventilator asynchrony dataset, and two electrocardiogram datasets, finding that the performance was either equal or superior to the state-of-the art on all three.