Infection is a global public health issue with high incidence and mortality. An international study conducted in 22 countries found that approximately 17% of patients who visited the emergency department (ED) presented with infections [1]. Infectious diseases constitute 5.4% of overall mortality in the United States and 26.1% of overall mortality worldwide [2,3]. Infection may progress to sepsis, a state of life-threatening organ dysfunction, which has a 90-day mortality rate of 32% [4,5].

Accordingly, early identification of serious infection in ED patients is crucial for rapid administration of antibiotics [6]. Delays in initiating appropriate antibiotics are associated with increased mortality in critically ill patients with infection [7]. However, inappropriate overuse of antibiotics results in antimicrobial resistance [6,8]. Hence, biomarkers or score-based models to identify infection or sepsis have previously been investigated [[9], [10], [11]].

Machine learning has recently been adopted to develop infection or sepsis prediction models, with excellent performance [[12], [13], [14], [15]]. Most models utilize a large number of features that are unavailable in the early stages of ED stay (such as continuous or repeated vital sign values, laboratory test results, and medications), thereby limiting their usefulness as a screening tool in the ED. [[16], [17], [18]] A new model that uses natural language processing (NLP) to extract information from ED clinical notes may help to overcome the limited amount of available information for early identification of serious infection in the ED.

A large proportion of patients visiting the ED, particularly the elderly, present without fever [19,20]. Identifying serious infection in these patients is critical to prevent delays in antibiotic administration and reduce adverse patient outcomes [21]. Therefore, the aim of this study was to develop and externally validate neural network and NLP-based models for identification of suspected serious infections in ED patients afebrile at initial presentation. We hypothesized that incorporating information extracted from text mining of ED physician notes would provide added value to a model that only included structured variables.