Machine learning prediction model for delirium after heart valve replacement with cardiopulmonary bypass: A large-scale cohort study

Comprehensive Summary

In this retrospective study, researchers developed and analyzed a traditional nomogram model and several machine learning (ML) models to predict postoperative delirium (POD) following heart valve replacement with cardiopulmonary bypass. POD was assessed using the CAM-ICU, with an incidence of 25.1% among the adult patient cohort (n = 1067) collected from First Affiliated Hospital of Xi’an Jiaotong University. The nomogram model (AUC = 0.754) identified nine peri-operative indicators of POD, including postoperative awakening time (PAT), postoperative mechanical ventilation time (PMVT), and postoperative re-intubation (POR). Out of the four ML models (DT, RF, SVM, and XGBoost) developed, the RF model had the highest predictive ability (AUROC = 0.854), and maintained its relatively high performance in an external validation (AUROC = 0.793). Researchers employed SHAP to identify the most significant predictors of POD; PAT, PMVT, and POR were also identified as critical factors by the RF model.

Outcomes and Implications

POD following heart valve replacement with cardiopulmonary bypass is associated with increased mortality and longer hospitalization. Identification of high-risk patients can help clinicians implement measures to help decrease preoperative risk, take more careful precautions inter-operatively, and closely monitor postoperative patients, all while preparing to enact targeted interventions if necessary. While the traditional method could create a practical and intuitive nomogram score to indicate high risk patients, the RF tree could identify nonlinear relationships and displayed a higher predictive ability. Thus, there is high value in further developing ML models with multi-center validation and taking steps to begin implementing these tools into a clinical workflow.