Introduction: The modified Duke criteria are recommended by current ESC guidelines as diagnostic algorithm in suspected infective endocarditis (IE). Given the categorization in major and minor criteria, it enables easy application into clinical routine, but may not optimally utilize the underlying information from each individual patient. In contrast, detailed statistical evaluation of multiple characteristics, especially when applied by artificial intelligence, report improved prediction of various cardiovascular diseases over conventional clinical strategies. We tested the hypothesis that neuronal networks and conventional logistic regression analysis would provide improved prediction of IE as compared to the modified Duke score.

Methods: This analysis is a post-hoc evaluation of the prospective observational PRO-ENDOCARDITIS study, conducted at the West German Heart and Vascular center. Between December 2017 and May 2019 a total of 261 patients referred to transesophageal echocardiography (TEE) with suspected IE were included. Duke criteria and clinical characteristics were prospectively collected. TEE imaging was evaluated by a blinded cardiologist at a central core-lab. IE as primary endpoint was adjudicated by an independent clinical endpoint committee. The database was divided into a training (70%) and validation cohort (30%). We compared the value of the Duke score, neuronal networks and logistic regression analysis for prediction of the primary endpoint. 

Results: Mean age of the cohort was 60.1 ± 16.1 years, 37.2% were female. In 47 cases, IE was present. As the current standard, the modified Duke score achieved an AUC of 0.863 in the training and 0.913 within the validation cohort. Both the logistic regression and the neural network exceeded the predictive value within the training and validation cohort (training cohort: 0.992 and 0.986; validation cohort: 0.964, 0.957; for logistic regression and neuronal networks, respectively, figure 1). Without the use of TEE, the remaining Duke criteria only poorly predicted IE (training cohort: 0.771, 0.951 and 0.938; validation cohort: 0.835, 0.862 and 0.780, for Duke score, logistic regression and neuronal networks, respectively).

Discussion: Logistic regression analysis and neuronal networks provide improved prediction of IE as compared to the clinically established Duke score. Further studies on larger databases are needed to confirm our results and provide algorithms for clinical routine.