To evaluate multiple machine learning algorithms for predicting in-hospital mortality among patients with acute left heart failure (ALHF), and to construct and externally validate a risk prediction model based on the optimal algorithm.

A total of 15 983 patients from the Medical Information Mart for Intensive Care-Ⅳ 3.1 database (training cohort) and 14 428 patients from the eICU database (external validation cohort) diagnosed with ALHF were included in this study. The outcome variable was in-hospital mortality among patients with ALHF. Machine learning analyses were performed using logistic regression, neural networks, stochastic gradient descent, scorecard modeling, adaptive boosting, k-nearest neighbors, classification and regression tree, and random forest algorithms. Univariate and multivariate logistic regression analyses were conducted in the training cohort to identify influencing factors associated with in-hospital mortality in patients with ALHF. Independent risk factors with an odds ratio (OR) ≥ 1.2 identified from the multivariate logistic regression were incorporated into a nomogram. The nomogram model was further validated using the external validation cohort. The predictive performance of the model was assessed by receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA).

Logistic regression demonstrated superior overall performance among all models, achieving the highest area under the curve (AUC) of 0.749, F1 score of 0.322, and Matthews correlation coefficient of 0.260. Multivariate logistic regression in the training cohort identified vasopressin, dopamine, malignant tumor, liver failure, endotracheal intubation, norepinephrine, epinephrine, mechanical ventilation, and acute myocardial infarction as independent risk factors with the OR ≥ 1.2. The nomogram constructed from these factors exhibited strong predictive performance with an AUC of 0.768 in the training cohort and 0.770 in the external validation cohort. Predicted probabilities closely matched observed outcomes in both cohorts, with calibration curves approaching the ideal reference line, indicating a good fit. DCA showed that within a high-risk threshold range of 0.13-0.92 in the training cohort, the model's net benefits significantly exceeded those of the "intervene-all" and "intervene-none" strategies, demonstrating robust clinical utility. Similar findings were observed in the external validation cohort within a high-risk threshold range of 0.13-0.93.

Among multiple machine learning algorithms, logistic regression provides the optimal prediction of in-hospital mortality risk for ALHF patients. The constructed nomogram incorporating key risk factors demonstrates favorable predictive accuracy and external validity, offering strong support for clinical identification of high-risk patients and optimization of therapeutic decision-making.