Machine Learning Superior to Logistic Risk Scores for Predicting Mortality Risk After TAVI

Machine learning was found to be superior to logistic risk scores in predicting intrahospital all-cause mortality after transcatheter aortic valve implantation (TAVI), according to study results published in Clinical Research in Cardiology.

Current strategies for identifying patients eligible for TAVI rely on risk assessment tools such as the Society of Thoracic Surgeon’s Risk Score (STS score). The predictive power of these tools is poor, and improved options for risk stratification of TAVI patients are needed.

In this retrospective analysis of data from 451 patients, investigators aimed to evaluate whether machine learning models could be used to predict clinical outcomes for patients after TAVI. A total of 83 features, including patient demographics, comorbidities, laboratory data, electro- and echocardiogram findings, and computed tomography (CT) results, were used to train and test the predictive models. Neural network, support vector machine, and random forest machine learning algorithms were used. The performance of each model was evaluated using 5-fold nested cross-validation and subsequently compared with standard risk assessment tools.

The characteristics that were found to be the most important in predicting all-cause mortality included New York Heart Association functional class, troponin T values, baseline and peak creatinine values, peak leucocyte count, left atrium diameter, fever after TAVI, pericardial effusion after TAVI, peak C-reactive protein levels, and female gender. Echocardiographic features of note included septal thickness, tricuspid annular plane systolic excursion, tissue Doppler e/e’ ratio, and systolic pulmonary artery pressure. Calcification severity of the ascending aorta and distance between the right coronary artery and annular plane on CT were also most predictive of all-cause mortality.

The random forest and support vector machine models had better capability of predicting all-cause mortality than the STS (P =.00007 and P =.037, respectively) and STS/American College of Cardiology Transcatheter Aortic Valve Replacement (STS/ACC TAVR) scores. The corresponding areas under the curve (AUC) for the random forest and support vector machine models were 0.81 and 0.82, respectively. The AUC for the neural network model, which was not superior to STS or STS/ACC TAVR scores, was 0.72.

After training the models on input data, the AUCs were improved to 0.97, 0.94, and 0.96 for the random forest, support vector machine, and neural network models, respectively. When using all available features, all models were significantly improved compared with using baseline characteristics alone.

The predictive performance of all models for stroke, major vascular complications, new pacemaker implantation, and paravalvular leakage was poor.

Study limitations include the relatively small sample population derived from a single center which may limit the generalizability of the results.

“Machine learning-derived prediction models seem to improve patient selection considering the outcome ‘intrahospital all-cause mortality’ compared [with] older risk scores,” the study authors concluded. “[T]he development of machine learning models using both structured and imaging data could further enhance the [artificial intelligence’s]’s potential for risk assessment.”

Disclosures: Dr Meder reported affiliations with Fleischhaker GmbH & Company KG.

Reference

Gomes B, Pilz M, Reich C, et al. Machine learning-based risk prediction of intrahospital clinical outcomes in patients undergoing TAVI [published online June 24, 2020]. Clin Res Cardiol. doi:10.1007/s00392-020-01691-0