The study team enrolled patients with critical pulmonary embolism from two multicenter critical-care databases. One center provided data for model development and internal validation; a second center supplied an external validation cohort. The two datasets contained 2,187 and 2,183 patients respectively. In the training cohort, 1,725 patients (78.88%) experienced acute kidney injury; the external validation cohort showed 1,185 cases (54.28%). These incidence figures framed subsequent model calibration and interpretation.
Investigators used a staged approach to variable selection. They applied least absolute shrinkage and selection operator (LASSO) regression followed by multivariate logistic analysis to reduce dimensionality and identify independent predictors. Eight machine learning algorithms were then developed from the selected features. The abstract lists the final predictors that entered the modelling pipeline: hypertension, heart failure, pneumonia, sepsis, ventilation, respiratory rate, low-density lipoprotein cholesterol, partial pressure of oxygen, lactate, and the Oxford Acute Severity of Illness Score. The source did not provide granular details of hyperparameter tuning, imputation strategies, or exact variable preprocessing steps in the abstract.
Model performance was assessed across three domains. Discrimination was measured with receiver operating characteristic analysis; calibration was examined with calibration curves; clinical utility was evaluated using decision curve analysis. Across training, internal validation, and external validation cohorts the extreme gradient boosting model delivered the most stable results. Reported area under curve (AUC) values for that model were 0.825 in the training set, 0.771 in internal validation, and 0.757 in external validation. The abstract states the model demonstrated consistent calibration and clinical benefit by decision curve analysis, but it does not include calibration slope, Brier scores, confidence intervals for AUCs, or detailed decision-curve thresholds in the abstract.
The study team quantified variable contributions using Shapley Additive eXplanation (SHAP) analysis. SHAP outputs indicated that the Oxford Acute Severity of Illness Score, low-density lipoprotein cholesterol, and partial pressure of oxygen were among the largest contributors to predicted AKI risk in the optimal model. The authors reported these features as high-impact drivers of model output in the abstract. The source did not provide full SHAP summary plots, feature value ranges, or interactions within the abstract text.
Investigators compared their machine-learning model against the pulmonary embolism severity index (PESI) score. Using net reclassification improvement (NRI) and integrated discrimination improvement (IDI) metrics, the proposed model demonstrated improved predictive performance relative to PESI. The abstract does not present numerical NRI or IDI values, nor does it specify whether comparisons were stratified by subgroups such as age or baseline renal function. The report therefore indicates superior performance but omits granular comparative statistics in the abstract.
A web-based calculator was deployed to translate the model into a clinician-facing tool. The authors described the calculator as a means to identify high-risk cohorts and support early, individualized management decisions. The abstract does not report implementation details such as the calculator's URL, user interface, input requirements, data security measures, or prospective usability testing. The intended clinical role—screening at ICU admission or guiding specific interventions—was described generally rather than specified with workflow integration steps.
The abstract lists multicenter development and external validation, but it leaves several questions open. It does not report confidence intervals for AUCs, specific sample selection criteria, or the timing and definition of AKI used for endpoint ascertainment within the abstract. Details about missing data handling, potential overfitting safeguards, and whether the web calculator was prospectively tested are not provided in the abstract. The source did note that the authors declared no funding or conflicts of interest. Further reading of the full text would be required to assess patient-level inclusion/exclusion criteria, calibration statistics, and whether model performance varies across clinically relevant subgroups.