A machine-learning model that helps medical institutions predict the mortality risk for cardiac surgery patients has been developed by researchers at Mount Sinai Health System.

The researchers at the New York hospital network said they saw a significant performance advantage over current population-derived models.

The data-driven algorithm, built on troves of electronic health records (EHRs), is the first institution-specific model for assessing a cardiac patient’s risk prior to surgery, thus allowing health care providers to pursue the best course of action for that individual.

The team’s work was described in a study published this month in The Journal of Thoracic & Cardiovascular Surgery (JTCVS) Open.

“The standard-of-care risk models used today are limited by their applicability to specific types of surgeries, leaving out significant numbers of patients undergoing complex or combination procedures for which no models exist,” said senior author Ravi Iyengar, director of the Mount Sinai Institute for Systems Biomedicine. “Our team rigorously combined electronic health record data and machine-learning methods to demonstrate for the first time how individual institutions can build their own risk models for post-cardiac surgery mortality.”

Prediction models based on machine-learning algorithms have been generated across diverse fields of medicine, and some have shown improved results over their standard-of-care counterparts. In cardiac surgery, the Society of Thoracic Surgeons (STS) risk scores are considered the gold standard, and are routinely used to assess a cardiac surgery patient’s procedural risk. While they continue to provide important benchmarks for hospitals to evaluate and improve their performance, they are derived from population-level data and therefore may fail to predict risk accurately for specific patients with complicated pathologies who require tailored preoperative evaluations and complex surgeries.

Cardiovascular surgeons and data scientists at the Mount Sinai Hospital, supervised by co-senior author Gaurav Pandey, associate professor at Icahn Mount Sinai, hypothesised that machine-learning models using EHR data from their own institution could be effective. So they created a rigorous machine-learning framework using routinely collected EHR data to develop a risk prediction model for postsurgical mortality that is personalised to the patient and specific to the hospital, implicitly incorporating important information about Mount Sinai’s patient population, such as demographics, socioeconomic factors and health characteristics.

This is in contrast to population-derived models such as STS, which are based on data from diverse health systems in different parts of the USA.

Further driving the performance of this method was an open-source prediction algorithm known as XGBoost, which builds an ensemble of decision trees by progressively focusing on harder-to-predict subsets of training data.

Leading the study was co-corresponding author Aaron Weiss, formerly a cardiothoracic surgery resident at Mount Sinai Hospital who is now at Cleveland Clinic. Arjun Yadaw, a senior data scientist at the National Center for Advancing Translational Science, also co-led the work.

These researchers used XGBoost to model 6392 cardiac surgeries performed at Mount Sinai Hospital from 2011 to 2016, including heart valve procedures; coronary artery bypass graft; aortic resection, replacement or anastomosis; and reoperative cardiac surgeries, which have been shown to increase mortality risk appreciably. The team then compared the performance of its model with STS models for the same patient sets.

The study showed that the XGBoost model outperformed STS risk scores for mortality in all commonly conducted categories of cardiac surgery for which STS scores were designed. Prediction performance of the XGBoost model across all surgery types was also high, demonstrating the potential of machine learning and EHR data for building effective institution-specific models.

“Accurate prediction of postsurgical mortality is critical to ensure the best outcomes for cardiac surgery patients, and our study shows that institution-specific models may be preferable to the clinical standard based on population data,” said Pandey. “Just as importantly, we’ve demonstrated that it’s practical for health care institutions to develop their own predictive models through sophisticated machine-learning algorithms to replace or complement the established STS template.”

This study was funded by grants from the National Institutes of Health.