Machine learning-based models for predicting mortality and acute kidney injury in critical pulmonary embolism

OBJECTIVES: We aimed to use machine learning (ML) algorithms to risk stratify the prognosis of critical pulmonary embolism (PE). MATERIAL AND METHODS: In total, 1229 patients were obtained from MIMIC-IV database. Main outcomes were set as all-cause mortality within 30 days. Logistic regression (LR) and simplified eXtreme gradient boosting (XGBoost) were applied for model constructions. We chose the final models based on their matching degree with data. To simplify the model and increase its usefulness, finally simplified models were built based on the most important 8 variables. Discrimination and calibration were exploited to evaluate the prediction ability. We stratified the risk groups based on risk estimate deciles. RESULTS: The simplified XGB model performed better in model discrimination, which AUC were 0.82 (95% CI: 0.78-0.87) in the validation cohort, compared with the AUC of simplified LR model (0.75 95% CI: 0.69-0.80). And XGB performed better than sPESI in the validation cohort. A new risk-classification based on XGB could accurately predict low-risk of mortality, and had high consistency with acknowledged risk scores. CONCLUSIONS: ML models can accurately predict the 30-day mortality of critical PE patients, which could further be used to reduce the burden of ICU stay, decrease the mortality and improve the quality of life for critical PE patients.