Promising Signs in PCI for the Expanding World of Big Data

Patrice Wendling

July 02, 2019

New data support the broadening implementation of artificial intelligence in medicine, suggesting that machine learning may outperform standard statistical methods to forecast patient prognosis after percutaneous coronary intervention (PCI).

A machine-learning algorithm was more predictive and discriminative than logistic regression models, trained on the same registry, at identifying patients at risk for 180-day cardiovascular (CV) mortality and 30-day congestive heart failure (CHF) rehospitalization. But it offered no advantage regarding in-hospital mortality among the more than 11,000 patients analyzed.

"You hear a lot about using these tools in medicine these days, and I think doing some proof-of-concept application on how they may improve things or in what particular areas they may improve things is important," cofirst author Conor Senecal, MD, Mayo Clinic College of Medicine, Rochester, Minnesota, told | Medscape Cardiology.

"Obviously there were some differences in our short-term versus our long-term models where the machine learning really outperformed the logistic regression, and I think that speaks to trying to help find a place where these things might be the most helpful."

The study was published online June 19 in JACC Cardiovascular Interventions.

Although there are many published traditional risk-prediction models for PCI, Senecal explained, uptake is limited in the clinical arena. Many models include the use of procedural information or narrowly defined or binary outcomes and require angiographic variables for risk prediction.

"What machine learning can do is leverage relationships that don't have a linear component to them and include a wider variety of patient variables, such as marital status, insurance, and comorbidities," he said.

To address these issues, the team, coled by Senecal and Chad Zack, MD, MS, Penn State Hershey Medical Center, Pennsylvania, evaluated data from 11,709 patients who underwent 14,349 PCIs from January 2004 until December 2013 in the Mayo Clinic PCI registry.

Of the 508 distinct variables in the registry, 52 known at the time of admission were used in models for in-hospital mortality and 410 in the models for CHF readmission and long-term mortality.

An independently developed machine-learning algorithm (Medial Research; Kfar Malal, Israel) model was trained to estimate the time to each event, with eight-fold cross-validation used to estimate model performance before comparison against internally trained logistic regression models derived from the same Mayo registry.

The mean age of the population was 66.9 years, three-fourths were male, and nearly half were obese. The median survival time was 3.2 years for patients who died during the study period.

The machine-learning model identified a high-risk cohort representing 2% of the study population with an in-hospital mortality of 45.5% compared with a risk of only 2.1% in the general population.

Although the model had excellent discriminatory ability (area under the curve [AUC], 0.925), it was not significantly better than the reference logistic regression model (AUC, 0.925; P = .84), the authors report. Age, CHF, and shock on presentation were leading predictors of in-hospital death.

One reason for the finding is that in-hospital mortality is heavily dependent on a few specific variables, such as cardiogenic shock, that are directly relatable to the patient's coronary disease, observed Senecal.

Second, the number of predictors in the models was smaller because only about one-fifth of the variables were available at admission. "The opposite is true for the more long-term outcomes with increased variables for analysis and with these likely able to identify patients at risk of complications due to their overall health than secondary to their CAD alone," he said. 

For CV mortality at 180 days, machine learning identified 1% of patients who had a substantially higher risk than the general population (13% vs 1.3%), with discriminatory ability that was significantly better than the logistic regression model (AUC, 0.812; P < .001). Leading predictors for the outcome were age, creatinine, chronic obstructive pulmonary disease, and CHF history.

Machine learning also identified 1% of patients at increased risk for 30-day CHF rehospitalization (8.1% vs 0.7%), with excellent discriminatory ability that outperformed the standard regression model (AUC, 0.85; P =.003). Length of stay, antiarrhythmic medication, and inotrope use were leading predictors.

"In this issue of the journal, Zack et al foreshadow the day when artificial intelligence and related big data constructs have a prominent, if not predominant, role in healthcare," R. Jeffrey Westcott, MD, Swedish Heart and Vascular Institute, Seattle, and James E. Tcheng, MD, Duke University Health System, Durham, North Carolina, say in an accompanying editorial.

They note, however, that many of the data points collected on each patient are unique to the Mayo registry and it would require a substantial investment in personnel that is "unlikely to be replicated elsewhere." Converting this, retrospective analysis to a real-time operation also would require additional resources.

The editorialists highlight the prediction of sepsis in the ICU as a "powerful success story" in machine learning, but say the key to its success is the availability of high-quality discrete data. Electronic health record (EHR) systems do not promote the capture of well-defined high-quality data and, despite its many contributions in cardiology, the current National Cardiovascular Data Registry model is resource-intensive and "should be modernized to become a workflow-integrated, patient-centric, and EHR integrated data management system."

"Transforming healthcare — and more specifically, transforming the management of data within healthcare to enable artificial intelligence and its siblings — will require both foundational investment and culture change," write Westcott and Tcheng. They go on to note that equity funding for healthcare artificial intelligence was expected to exceed $2.4 billion last year, up 78% from 2017.

Senecal said their findings need to be prospectively validated and tested in-real time, and agree it would be labor-intensive to replicate the results using available datasets or with a machine-learning algorithm at another institution with different data-collection processes.

"Looking at the big picture moving forward, one of the goals would be for the electronic health record to do this behind the scenes and really be able to pull in the appropriate data and pop up saying, 'this is the risk-prediction model for this particular patient that you're looking at'," senior author Rajiv Gulati, MD, PhD, also from the Mayo Clinic, told | Medscape Cardiology.

"So it's really integration with the health record moving forward. And the advantage of machine learning is a continuously learning process, so it will get better and better the more data that are inputted into the algorithms down the line. So a continuously learning [algorithm] that interfaces seamlessly with the electronic health record would be the ultimate goal," he said.

"We recognize there is a way to go before that happens, but that would be one way this would be clinically relevant for practicing clinicians, absolutely."

Support for this study was provided by a National Institutes of Health CTSA grant from the National Center for Advancing Translational Sciences. Senecal, Gulati, Westcott, and Tcheng reported no relevant financial relationships.

JACC Cardiovasc Interv. Published online June 19, 2019. Abstract, Editorial

Follow Patrice Wendling on Twitter: @pwendl. For more from | Medscape Cardiology, join us on Twitter and Facebook.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: