A JACC: Cardiovascular Interventions study published this month suggests machine learning models are more predictive and discriminative than standard methods for identifying patients at the greatest risk of CV mortality and rehospitalization after percutaneous coronary intervention (PCI).
First author Chad J. Zack, MD, MS, and colleagues said in the journal that since PCI is such a common procedure for heart patients, it’s important health systems are able to flag those at the highest odds for complications. But today’s regression-based prediction models have been limited in their success, struggling with linearity and variable selection to accurately identify patients at the greatest risk for death or congestive heart failure (CHF) rehospitalization.
“Advances in machine learning have resulted in the creation of algorithms capable of analyzing large volumes of patient-level data to assist in disease diagnosis, risk assessment and clinical decision-making,” the authors wrote. “We sought to address the limitations of regression-based PCI risk models by using machine learning methods to derive discriminatory models that preemptively identify patient populations at risk for mortality and rehospitalization after PCI.”
Zack et al. analyzed 11,709 patients who’d undergone 14,349 total PCIs between January 2004 and December 2013 in the Mayo Clinic PCI registry. Using 52 demographic and clinical parameters available at index admission and 358 additional variables available at discharge, they predicted the odds of in-hospital mortality and CHF readmission among the study population.
For each event, the researchers trained a random forest regression model to estimate time-to-event. The predictive algorithm identified a high-risk cohort representing 2% of all patients who had an in-hospital mortality of 45.5% compared with a risk of 2.1% in the general population.
Older age, CHF and shock at presentation were the leading predictors for outcome, the authors reported. Random forest regression outperformed logistic regression for predicting 30-day CHF readmissions, achieving an area under the curve (AUC) of 0.90 compared to 0.85 in logistic regression models. Machine learning also outperformed traditional methods for predicting 180-day cardiovascular death, with AUCs of 0.88 and 0.81, respectively.
In a related editorial, R. Jeffrey Westcott, MD, and James E. Tcheng, MD, said Zack and colleagues’ findings support the idea that machine learning could outperform classical statistical approaches to risk prediction—but it’ll take some work to make it an industry standard.
“Transforming healthcare, and, more specifically, transforming the management of data within healthcare to enable AI and its siblings, requires foundational investment and culture change,” the editorialists wrote.
They said artificial intelligence and machine learning will undoubtedly become “increasingly important in clinical medicine” as we move forward, with equity funding for healthcare-related AI ventures topping $2.4 billion in 2018. That’s a 78% increase from 2017 alone.
“Machine learning has proven to be valuable and is therefore the future,” Westcott and Tcheng wrote. “Data warehouses and data lakes contain amazing amounts of structured and unstructured data that will change how medical research, drug and device trials, and device tracking are done. A collaborative effort is needed with EHR vendors, third-party vendors, professional societies and others to start meaningful standardized data collection and workflow redesign now.”