An interpretable machine-learning model for predicting postoperative recovery quality after cardiovascular surgery: development, validation, and clinical applicability
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objectives
Quality of recovery (QoR) following cardiovascular surgery represents a key patient-centered outcome closely related to complications, hospital stay, and resource utilization. This study aimed to develop and validate an interpretable machine-learning model for predicting early postoperative recovery quality after cardiovascular surgery and to derive clinically actionable risk stratification to guide perioperative management.
Methods
We retrospectively analyzed 581 adult patients who underwent cardiovascular surgery at the Affiliated Hospital of Yangzhou University between March 2021 and September 2025. The primary endpoint was poor recovery, defined as QoR-15 < 90 on postoperative day 3. Predictor variables included demographic, ASA classification, emergency status, cardiopulmonary bypass (CPB), preoperative lactate, surgical duration, rebeating strategy, and modified Frailty Index (mFI). Data were randomly split 7:3 into training and test sets, with the final 20% of patients used for temporal external validation. Six ML algorithms include logistic regression (LR), K-nearest neighbors (KNN), Extremely Randomized Trees (ExtraTrees), Support Vector Machines (SVMs), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost) were compared using 10-fold cross-validation and hyperparameter optimization. Model discrimination, calibration, and clinical utility were evaluated using AUC, calibration plots, the Hosmer-Lemeshow test, and decision curve analysis (DCA). Model interpretability was assessed with SHapley Additive exPlanations (SHAP), and risk thresholds were derived from DCA for practical clinical stratification.
Results
Among the 581 patients, 173 (29.8%) experienced poor recovery. The XGBoost model achieved the best overall performance (AUC = 0.982, accuracy = 0.974, Hosmer–Lemeshow p = 0.791) with excellent calibration and temporal validation (AUC = 0.997). SHAP analysis identified five key predictors of poor recovery: female sex, higher ASA grade, elevated preoperative lactate (>2 mmol/L), longer operative duration, and greater frailty (mFI ≥ 0.25). Risk thresholds derived from DCA defined three clinical tiers-low (<0.15), intermediate (0.15-0.40), and high (>0.40)-for tailored postoperative management.
Conclusions
An interpretable XGBoost model accurately predicted postoperative recovery quality after cardiovascular surgery using routinely collected clinical data. The model’s transparency enables identification of modifiable risk factors and supports personalized perioperative optimization. Multicenter prospective validation and integration into perioperative decision-support systems are warranted to enhance recovery-oriented, patient-centered outcomes.
