In the Journals

Machine-learning algorithm predicts outcomes after CRT in HF

Michael Gold, MD, PhD
Michael R. Gold

Clinical outcomes in patients with HF with intraventricular conduction delay and reduced left ventricular function were predicted using a machine-learning algorithm, which may better differentiate outcomes compared with current clinical discriminators, according to a study published in Circulation: Arrhythmia and Electrophysiology.

“Machine learning is a powerful, computational method that could allow for improved description of phenotypes and development of decision support tools to predict clinical outcomes and better inform shared decision-making with patients,” Matthew M. Kalscheur, MD, assistant professor of cardiovascular medicine at University of Wisconsin Hospital in Madison, and colleagues wrote.

Researchers analyzed data from 595 patients from the COMPANION trial with advanced HF with an LV ejection fraction less than 35% who received cardiac resynchronization therapy defibrillators.

The machine-learning model was developed with the random-forest algorithm. This model was compared with a combination of bundle branch block morphology and QRS duration.

The outcomes of interest were the composite of HF hospitalization or all-cause mortality or all-cause mortality alone.

During a median follow-up of 15.7 months, 105 deaths occurred.

Statistical significance was not achieved in the difference in the survival distribution using bundle branch block morphology and QRS duration (log-rank P = .08). Using the random-forest algorithm, significance was reached in the difference in survival distribution (P < .0001). There was an eightfold increase in mortality in quartile 4 (highest predicted probability of events) compared with quartile 1 (lowest predicted probability of events; HR = 7.96; 95% CI, 3.6-17.56).

Patients in quartile 4 had the worst outcomes predicted by the random-forest algorithm, but the best outcomes using bundle branch block morphology and QRS duration. These patients had significantly more instances of HF hospitalization (HR = 2.42; 95% CI, 1.09-5.34) and all-cause mortality (HR = 6.62; 95% CI, 1.55-28.3) compared with patients in quartile 1.

“As clinical data sets expand, application of machine-learning algorithms will lead to further improvements in precision cardiovascular medicine,” Kalscheur and colleagues wrote.

“Certainly, there are precedents for robust complex models that retain their value over long periods of the evolution of therapy, such as the Seattle Heart Failure model,” Daniel B. Cobb, MD, electrophysiology cardiology fellow at Medical University of South Carolina in Charleston, and Michael R. Gold, MD, PhD, professor, associate dean of interdisciplinary clinical programs and Michael E. Assey chair in cardiology at Medical University of South Carolina, wrote in a related editorial. “Hopefully, the present model also sustains its value despite the continued and rapid evolution of HF treatment and CRT.” – by Darlene Dobkowski

Disclosures: The authors, Cobb and Gold report no relevant financial disclosures.

Michael Gold, MD, PhD
Michael R. Gold

Clinical outcomes in patients with HF with intraventricular conduction delay and reduced left ventricular function were predicted using a machine-learning algorithm, which may better differentiate outcomes compared with current clinical discriminators, according to a study published in Circulation: Arrhythmia and Electrophysiology.

“Machine learning is a powerful, computational method that could allow for improved description of phenotypes and development of decision support tools to predict clinical outcomes and better inform shared decision-making with patients,” Matthew M. Kalscheur, MD, assistant professor of cardiovascular medicine at University of Wisconsin Hospital in Madison, and colleagues wrote.

Researchers analyzed data from 595 patients from the COMPANION trial with advanced HF with an LV ejection fraction less than 35% who received cardiac resynchronization therapy defibrillators.

The machine-learning model was developed with the random-forest algorithm. This model was compared with a combination of bundle branch block morphology and QRS duration.

The outcomes of interest were the composite of HF hospitalization or all-cause mortality or all-cause mortality alone.

During a median follow-up of 15.7 months, 105 deaths occurred.

Statistical significance was not achieved in the difference in the survival distribution using bundle branch block morphology and QRS duration (log-rank P = .08). Using the random-forest algorithm, significance was reached in the difference in survival distribution (P < .0001). There was an eightfold increase in mortality in quartile 4 (highest predicted probability of events) compared with quartile 1 (lowest predicted probability of events; HR = 7.96; 95% CI, 3.6-17.56).

Patients in quartile 4 had the worst outcomes predicted by the random-forest algorithm, but the best outcomes using bundle branch block morphology and QRS duration. These patients had significantly more instances of HF hospitalization (HR = 2.42; 95% CI, 1.09-5.34) and all-cause mortality (HR = 6.62; 95% CI, 1.55-28.3) compared with patients in quartile 1.

“As clinical data sets expand, application of machine-learning algorithms will lead to further improvements in precision cardiovascular medicine,” Kalscheur and colleagues wrote.

“Certainly, there are precedents for robust complex models that retain their value over long periods of the evolution of therapy, such as the Seattle Heart Failure model,” Daniel B. Cobb, MD, electrophysiology cardiology fellow at Medical University of South Carolina in Charleston, and Michael R. Gold, MD, PhD, professor, associate dean of interdisciplinary clinical programs and Michael E. Assey chair in cardiology at Medical University of South Carolina, wrote in a related editorial. “Hopefully, the present model also sustains its value despite the continued and rapid evolution of HF treatment and CRT.” – by Darlene Dobkowski

Disclosures: The authors, Cobb and Gold report no relevant financial disclosures.