Development of a Machine Learning-Based Prediction Model to Differentiate Infectious and Non-Infectious Diseases in Patients with Undiagnosed Fever: A Single Hospital-Based Retrospective Study

Background/Objectives: Fever can develop from several causes, including infectious diseases, noninfectious inflammatory diseases (NIID), malignancies, and other medical conditions. Although serum ferritin (SF) level can help differentiate infectious from non-infectious diseases, its discriminative ability (specificity) is far from satisfactory. The aim of this study was to develop a diagnostic prediction model to distinguish infectious diseases from other febrile illnesses using only common blood tests available on admission, in addition to SF level, in patients with undiagnosed fever. Methods: This single-center retrospective observational study included patients with fever of unidentified origin aged ≥18 years admitted to a Japanese acute care hospital between 1 January 2013, and 31 December 2022. They were divided into infectious and non-infectious disease groups based on their final diagnosis. Machine learning and multivariable logistic regression analysis were used to develop a model to differentiate infectious diseases from non-infectious diseases. Model performance was evaluated using area under the curve (AUC), shrinkage coefficient, and stratified likelihood ratio. Results: Among the 143 patients included, 73 had infectious diseases. A prediction model consisting of five factors—serum white blood cell count, neutrophil percentage, platelet count, lactate dehydrogenase level, and log-transformed SF level—was developed. The AUC of the model was 0.794 (95% confidence interval: 0.721–0.867) with a sensitivity of 77.1%, specificity of 68.5%, shrinkage coefficient of 0.876, and stratified likelihood ratio of 0.13–5.04. Conclusions: We developed a prediction model consisting of only five high-performing indicators, which would help differentiate infectious diseases from other fever causes early after admission.