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

Figures

• 

  Fig. 1. Study design and dataset partitioning. Schematic illustration of dataset construction and analysis workflow. The study cohort was divided into five non-overlapping patient populations: a development cohort, a tuning cohort, an internal test cohort from an academic medical center, an external test cohort from a community hospital, and the MIMIC-IV test cohort. Patient assignment to mutually exclusive datasets prevented data leakage between model development and validation phases

• 

  Fig. 2. Discriminative performance across test cohorts. Receiver operating characteristic (ROC) curves (a) and precision-recall curves (b) for the deep learning model across the internal test set (AUC 0.821, 95% CI 0.804–0.838), external test set (AUC 0.800, 0.782–0.818), and MIMIC-IV test set (AUC 0.756, 0.743–0.768). N = 22,205 (internal), 22,709 (external), and 66,937 (MIMIC-IV). Circle markers denote the Youden-optimal operating point; the recommended high-sensitivity rule-out operating points and the full threshold spectrum are provided in Supplementary Table S6

• 

  Fig. 3. Subgroup analysis across patient characteristics. Point estimates (with 95% confidence intervals) of model performance (AUC), shown as points with error bars, stratified by demographic and clinical characteristics: age categories, gender, Charlson comorbidity index, qSOFA score, and, in the MIMIC-IV cohort, race and ethnicity. Results are presented separately for the internal test, external test, and MIMIC-IV test cohorts

• 

  Fig. 4. Survival analysis by AI-ECG risk classification. Kaplan-Meier survival curves for 90-day all-cause mortality, stratified by AI-ECG risk categories: high-risk (prediction score > 1.625), medium-risk (0.827 to 1.625), and low-risk (< 0.827). Upper panels present mortality in the overall population; lower panels display mortality within BSI-positive patients. Adjusted hazard ratios (adjusted for age and sex) with 95% confidence intervals are shown for each risk category relative to the low-risk reference group

• 

  Fig. 5. Key ECG features for bloodstream infection prediction. Left panels show relative feature importance based on information gain derived from XGBoost models. R-squared values indicate the proportion of variance explained. Right panels provide comparative analysis of key ECG parameters between patients with and without bloodstream infection, demonstrating statistically significant differences