Machine learning model improves mortality risk prediction in heart surgery

Important points:

• Machine learning-based models predicted mortality after cardiac surgery more accurately than population-based models.
• Accuracy was consistent across the five surgical types.

Machine-learning-based models appear to better predict mortality risk in patients undergoing cardiac surgery compared to population-derived models, researchers reported.

“Standard-of-care risk models in use today are limited by their applicability to certain types of surgery, excluding a significant number of patients undergoing complex or combination surgeries for which models do not exist. .” Dr. Ravi Iyengar, Dorothy H. Rosenstill and Lewis Rosenstill, Professor of Pharmacology at the Icahn School of Medicine at Mount Sinai and Director of the Mount Sinai Systems Biomedical Institute, said in a press release. “Our team is the first to demonstrate how an individual hospital can build its own risk model for post-cardiac surgery mortality using a rigorous combination of electronic medical record data and machine learning techniques.”

Ravi Iyengar

Based on a cohort of patients who underwent cardiac surgery between 2011 and 2016, the researchers developed a “data-driven site-specific machine learning-based model inferred from multimodal electronic medical records.” We developed and compared performance with the Society of Thoracic Surgeons (STS). )model. ”

This cohort included 6,392 patients described by 4,016 features and randomly assigned to a training/development cohort (75%) or a test/evaluation cohort (25%).

The researchers collected data from a diverse population on a regular basis to build a hospital-specific postoperative mortality risk prediction model tailored to the patient, as opposed to the STS population-based model. developed a machine learning framework using the EHR data collected. from different locations.

Researchers trained four machine learning classification algorithms based on imputed data.

According to the researchers, the best performing of these predictors was the eXtreme Gradient Boosting (XGBoost) algorithm.

Iyengar et al. showed that XGBoost performed well (F-value = 0.775; precision = 0.756; recall = 0.795; precision = 0.986; area under the receiver operating characteristic curve = 0.978; area = 0.804), outperforming XGBoost. STS population-based model for evaluating five different surgical procedures.

Gaurav Pandey

“Accurately predicting postoperative mortality is critical to ensuring the best outcome for cardiac surgery patients, and our study favors site-specific models rather than clinical standards based on population data.” It shows that it has potential.” Dr. Gauraf Pandey, Associate Professor of Genetics and Genomic Sciences at Icahn Mount Sinai said in a release. “Equally important, we have demonstrated that it is practical for healthcare organizations to develop their own predictive models through sophisticated machine learning algorithms to replace or complement established STS templates. .”

reference: