A Novel Method for Cardiac Risk Assessment using Neural Networks

Abstract

Introduction: While metrics of cardiac risk factors abound, they largely suffer from a common issue - reductionism.

In this abstract, we propose a naive big-data approach to risk factor estimation. Utilizing a linguistic neural network trained on massive health insurance dataset, we attempt to learn the "hidden grammar" of heart disease. This tool is intended as a hypothesis generator, and a risk-analysis tool for population health.

Hypothesis: Human bodies generate a certain amount of chaos (entropy) during our lifetime, as reflected in the time-series data captured in EMR systems or insurance claim data. We hypothesize that, by studying this data stream, we can distinguish between (a) relatively healthy patients, (b) patients that have suffered an acute, but treatable event, (c) patients with a managed chronic disease, and (d) patients at the end-of-life stage.

We further hypothesize that, by training machine learning algorithms on entropy time-series data, we can predict certain types of health events with a lead time allowing for intervention and prevention.

Methods: Utilizing Medicare Public Release file, we have compiled patient trajectories for n=600,000 patients who have been admitted to hospice due to a cardiac event, or suffered a cardiac event (MI or stroke) in the span of 36 months. By combining diagnostic (ICD-10), procedure (HCPCS) and prescription codes, we have built a set of “patient trajectories” corresponding to all events leading up to the acute event or hospice admission.

An LSTM neural network was trained on 80% of these trajectories, learning “latent semantic graph” of cardiac treatment, as well as learning to reject irrelevant data. The remaining 20% of patient trajectories were used as a test set, with the neural network attempting to predict time of hospice admission, stroke or MI.

Results: The neural network predicts the likelihood of hospice admission within 90 days with the f-score of 0.87%, and likelihood of MI or stroke within 90 days with f-score of 0.72%.

Conclusions: While the current f-scores are insufficient for clinical use, this method is capable of long-term forecasting of cardiac risk on population level, and will evolve over time to allow for clinical interventions.

This is a “clinically naive” technology that is entirely driven by the data. This is both an advantage and a disadvantage. On one hand, this tool can potentially alert clinicians and researchers to previously undetected risk factors; on the other hand it has likely learned existing clinical biases and may perpetuate bad decisions.

We believe that this technology has a lot of potential, as well as many potential pitfalls. We welcome your feedback as the ideas and tools behind medical AI evolve.

Keywords: cardiac risk, heart attack, neural network, insurance data