Using Artificial Intelligence (AI)
to Predict Arrhythmia and Sudden Death

By Adrienne Mueller, PhD
December 9, 2020

If you’ve had a heart attack, you may think the worst is behind you, but heart attack survivors can also develop a condition called ventricular arrhythmia which, in the worst case, can lead to sudden cardiac death. Currently, it is very hard to predict whether a given heart attack survivor is likely to develop ventricular arrhythmia. Traditional clinical markers are, at best, only 64% accurate as predictors for developing the condition. Random chance alone is 50% accurate. Notably, lots of potential causes have been shown at the cell level, but these cannot currently be used to predict sudden death in patients.

In their study recently published in Circulation Research, co-first authors Albert J. Rogers, MD and Anojan Selvalingam and senior author Sanjiv Narayan, MD used a machine learning (AI) approach based on cell-level data to improve our ability to predict which heart attack survivors are more or less likely to develop sudden death. “We hypothesized that the shape of signals recorded in patients reflect detailed cell types that AI can recognize to predict sudden death," commented Dr. Narayan. Their team collected almost 6000 electrical recordings (specifically monophasic action potentials) from the hearts of 42 heart attack survivors who had never had ventricular arrhythmias, and trained a machine learning model (support vector machine) to identify which patterns in the recordings were predictive of either 1) developing sudden death from ventricular arrhythmia or 2) dying within three years. Their machine learning-derived “computational phenotype” vastly outperforms existing clinical marker models – predicting the development of ventricular arrhythmia with 90% accuracy and fatality with 91% accuracy.

Figure: Machine learning-computational phenotypes identified differences in the shape of heart electrical signals (monophasic action potentials) which predict sudden death, and identify changes in calcium fluxes as mechanism.

Because machine learning is often considered a black box that cannot be interpreted, the investigators set out to closely study the computational phenotype to shed light on the mechanism underlying the difference in susceptibility. They found that individuals more susceptible to ventricular arrythmia had hearts that exhibited abnormal calcium fluxes. This finding points the way for future studies to explore the cellular mechanisms causing these abnormal calcium fluxes, and work towards better means of preventing and treating the condition.

By identifying a computational phenotype, the investigators have found a superior marker for the likelihood of developing ventricular arrhythmia. In the future, computational phenotypes that reflect specific cellular mechanisms identified by machine learning to predict a defined outcome may allow specific subsets of patients to be more closely monitored, receive personalized therapy, and ultimately have better outcomes. The success of this study suggests that machine learning-derived computational phenotypes could be considered for the diagnosis of other cardiac and non-cardiac disorders.

Stanford Cardiovascular Institute-affiliated authors who contributed to this work include Albert J. Rogers, Anojan Selvalingam, Mahmood I. Alhusseini, Tina Baykaner, Paul Clopton, Paul J. Wang, and Sanjiv M. Narayan.

Dr. Albert J. Rogers

Dr. Anojan Selvalingam

Dr. Sanjiv Narayan