Use of machine learning algorithms to predict life-threatening ventricular arrhythmia in sepsis

AimsLife-threatening ventricular arrhythmias (LTVAs) are common manifestations of sepsis. The majority of sepsis patients with LTVA are unresponsive to initial standard treatment and thus have a poor prognosis. There are very limited studies focusing on the early identification of patients at high risk of LTVA in sepsis to perform optimal preventive treatment interventions. We aimed to develop a prediction model to predict LTVA in sepsis using machine learning (ML) approaches.Methods and resultsSix ML algorithms including CatBoost, LightGBM, and XGBoost were employed to perform the model fitting. The least absolute shrinkage and selection operator (LASSO) regression was used to identify key features. Methods of model evaluation involved in this study included area under the receiver operating characteristic curve (AUROC), for model discrimination, calibration curve, and Brier score, for model calibration. Finally, we validated the prediction model both internally and externally. A total of 27 139 patients with sepsis were identified in this study, 1136 (4.2%) suffered from LTVA during hospitalization. We screened out 10 key features from the initial 54 variables via LASSO regression to improve the practicability of the model. CatBoost showed the best prediction performance among the six ML algorithms, with excellent discrimination (AUROC = 0.874) and calibration (Brier score = 0.157). The remarkable performance of the model was presented in the external validation cohort (n = 9492), with an AUROC of 0.836, suggesting certain generalizability of the model. Finally, a nomogram with risk classification of LTVA was shown in this study.ConclusionWe established and validated a machine leaning-based prediction model, which was conducive to early identification of high-risk LTVA patients in sepsis, thus appropriate methods could be conducted to improve outcomes. Graphical AbstractOverview of the study design. A total of 27 139 patients with sepsis were randomly divided into the training set and internal validation set at a ratio of 8:2. Several machine learning algorithms including CatBoost, XGBoost, and logistic regression were used to perform the model fitting. After key feature selection, hyper-parameter optimization was implemented to modify the model. Finally, 9492 sepsis patients from another database were involved to conduct external validation.