Concordance Between Dispatch Suspicion, On‑Scene Phenotype, and Time Sensitive Triage in Prehospital Infectious Presentations: A Retrospective Machine‑Learning Study

Background: Early risk stratification of infectious presentations during emergency calls is challenging; symptom-based dispatch frequently diverges from physiological severity observed on scene. The concordance between the Emergency Medical Call Centres (EMCC) presenting problem (“Fever/Infection”), the first on-scene Emergency Symptoms and Signs (ESS) label, and high-risk triage (RETTS Red/Orange), where examined and evaluated whether routinely captured metadata can stratify risk in this cohort. Objective: To explore concordance among the EMCC reason for call ("Fever/Infection"), the first on‑scene ESS label, and HRTS (RETTS Red/Orange) at first EMS contact, and to evaluate whether routinely captured metadata can stratify risk in this cohort. Methods: We performed a population‑based retrospective study of EMS assignments in Region Stockholm (2017–2022). After prespecified exclusions, 34 779 assignments initially categorized as "Fever/Infection" at EMCC were analysed. The primary endpoint was high‑risk triage (HRTS = RETTS Red/Orange) at first EMS contact. Machine‑learning models (gradient boosting [primary], random forest, logistic regression, and ensembles) used routinely recorded variables: age, sex, call hour, month, response time, ambient temperature, a workload proxy (mean response time in the current hour), and the first on‑scene ESS category. Discrimination (AUC), KS statistic, misclassification, and partial‑dependence (PD) analyses were reported on a held‑out test set. Results: Infection‑coded ESS labels were present in 64.0% of assignments. Within infection‑coded ESS, 71.5% were triaged HRTS versus 44.3% for other ESS categories. Gradient boosting performed best (AUC 0.724). Variable‑importance ranked ESS keywords, response time, and age as most influential. PD analyses showed a near‑monotonic decline in predicted HRTS probability with increasing response time, a threshold‑like rise beyond ~60 years, modest circadian/seasonal patterns, and non‑linear temperature effects. Because ESS contributes to RETTS, ESS–HRTS alignment is interpreted as concordance rather than independent prediction. Conclusions: Routinely recorded dispatch and early on-scene metadata can moderately stratify severity in infectious EMS calls and make visible where dispatch suspicion, on-scene phenotype, and triage align. While not ready for operational deployment, these findings support secondary triage use cases (e.g., prioritizing delayed older patients) and motivate prospective evaluation and augmentation with richer signals.