Development and validation of an interpretable machine learning model for predicting the risk of non-cardiac surgery postoperative heart failure: a multicenter study

Abstract

Background: This study developed a machine learning model to predict postoperative heart failure (HF) risk in non-cardiac surgery patients. Methods: Using data from 489 patients (109 HF cases, 380 controls), the dataset was split 8:2 into training and testing sets, with under-sampling for class imbalance. Eight algorithms were evaluated, with random forest (RF) performing best. Results: The RF model achieved AUROCs of 0.919 (training) and 0.923 (testing), validated externally (AUC = 0.878). SHAP analysis identified key predictors: age, neutrophil-to-lymphocyte ratio, blood glucose, INR, pulse and serum creatinine (positively associated); serum albumin, MCHC, eGFR and diastolic blood pressure (negatively associated). A web-based tool was developed for clinical use. Conclusion: The model integrates 10 clinical variables reflecting age, inflammation, renal dysfunction, and hemodynamic instability, enabling preoperative risk stratification and guiding targeted interventions to improve perioperative outcomes.