AI-based clinical prediction model for early infectious disease classification in the emergency department

Abstract

Background Diagnosing infections remains challenging. Clinicians rely on scoring systems and experience, but artificial intelligence (AI) is used to support decision-making by integrating clinical data. However, most AI models focus on predicting adverse outcomes (e.g., ICU admission or sepsis) rather than differentiating between infection types. Objective To develop predictive models and evaluate their discriminatory performance for classifying infected patients into pneumonia, urinary tract infection (UTI), and other infections, and to assess the incremental value of adding clinical tests to routine variables. Methods Six Random Forest models were developed on datasets sharing a set of standard variables but differing in the added clinical variables. The models predicted infection type across three categories using a one-vs-rest framework. Data were divided into 70% for model training and 30% for model testing. The decision threshold was set with sensitivity fixed at 0.8, and specificity, negative predictive value, positive predictive value, and accuracy were reported. Variable importance was assessed, and the top five predictors per model were identified. Results The model including standard variables and urine data achieved the highest AUC for UTI, whereas the total model achieved the highest AUC for pneumonia. Overall, urine and total models demonstrated the strongest performance. Key predictors included positive chest X-ray findings, abnormal auscultation, suspected pulmonary disease on imaging, neutrophil count, weight, urine culture, and a leukocyte-positive dipstick. Conclusion Machine learning models can potentially improve infection classification at ED admission when routine data are supplemented with relevant diagnostics. Optimal combinations of clinically relevant data markedly enhance diagnostic performance.