Deep learning analysis of 12-lead electrocardiograms for bloodstream infection prediction: a multi-center validation study

BACKGROUND: Prompt diagnosis of bloodstream infections (BSIs) is critical for antimicrobial stewardship but hindered by blood culture delays of 48 h or more. We developed and validated a deep learning model to predict BSI risk from standard 12-lead electrocardiograms (ECGs). METHODS: This study adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines. In this retrospective, multi-center study, we trained a deep learning model to predict clinically significant BSIs (excluding common contaminants) using ECGs from adult emergency department (ED) patients at an academic medical center in Taiwan (April 2012 to March 2022; total N = 163,826) and validated it in three independent cohorts: internal (n = 22,205), external (n = 22,709), and MIMIC-IV (n = 66,937). For the MIMIC-IV cohort, BSI was identified using the derived suspicion of infection framework with a positive blood culture flag. RESULTS: The model achieved AUCs of 0.821 (95% CI 0.804-0.838), 0.800 (0.782-0.818), and 0.756 (0.743-0.768) across the three test cohorts, respectively. At a high-sensitivity rule-out threshold the model classified 19-25% of patients as low risk (negative predictive value [NPV] 99.8%); at the Youden point, the sensitivity and specificity were 73.2% and 76.8% in the internal test set, with an NPV of 99.2%. The model showed superior discrimination compared with XGBoost and logistic regression baselines. Decision-curve analysis showed net benefit over heart rate alone and the default treat-all and treat-none strategies across the clinically relevant low-threshold range. Exploratory analyses indicated that the ECG-derived risk signal was elevated across Gram-positive, Gram-negative, and fungal infections; pathogen-specific prediction will require a dedicated model (Supplementary Figure S3). The primary model stratified 90-day mortality risk among infected patients (adjusted HR 3.43 [1.93-6.09] for high- vs. low-risk). CONCLUSIONS: Deep learning analysis of standard 12-lead ECGs obtained at ED triage can predict clinically significant BSI with discrimination superior to conventional ECG parameters. The model discriminated best (highest AUC) in patients with low clinical suspicion (qSOFA = 0), although its incremental value in this low-prevalence subgroup is modest. Prospective validation is needed before clinical implementation.