Prediction Models of Infective Endocarditis Usable Ahead of Performing Blood Cultures: A Narrative Review

Infective endocarditis (IE) often presents as a fever of unknown origin due to its extremely diverse clinical presentations, requiring diverse advanced medical equipment and tests to make a correct diagnosis. Whether a physician can suspect IE in a clinical setting is dependent on the physician's knowledge and experience. If IE is not suspected, antibiotics are administered without obtaining blood cultures, complicating the clinical course and prognosis. To avoid delayed diagnosis or entering the maze of diagnostic difficulties of IE cases, a prediction model to deduce IE likelihood can be used at an early stage after a patient's arrival at the hospital before blood culture examinations would be invaluable. In this study, we aimed to review the literature on such prediction models for IE diagnosis in existence, discussing their strengths and limitations. A narrative review was conducted by two researchers using PubMed. Comprehensive searches included the index terms "infective endocarditis" or "infectious endocarditis", coupled with "prediction model" or "prediction rule" or "predictive model". Five articles reporting one of the three prediction models were identified. The first model, developed for intravenous drug users (IDUs) admitted to the emergency departments of two to three hospitals showed a good area under the curve (AUC) of 0.8; however, the small sample size and overfitting of the model were a limit. The second model for inpatients in all departments of four hospitals showed an AUC of 0.783 with a shrinkage coefficient of 0.963, indicating high generalizability. Moreover, it featured the highest ease of use because it consisted of only five factors readily available in any hospital. The third model, developed for inpatients admitted to an emergency department at a single center, consisted of 12 factors and achieved the highest AUC (0.881). All models demonstrated fair to good AUC. The second model excelled in generalizability and ease of use, while the third model was superior in performance. To further improve the accuracy of each IE prediction, further high-level evidence studies, such as randomized controlled trials in multiple facilities, are mandatory.