Development and validation of a machine learning model to predict the vasopressor requirement of sepsis patients in the intensive care unit
Jisoo Lee, Jungyun Oh, Ho-Geol Ryu, Hyeonhoon Lee, Hong Yeul Lee, Jinwoo Lee, Hannah Lee, Kwangsoo Kim, Hyun-Lim Yang, Seung-Young OhObjective
The purpose of this study is to predict the required vasopressor doses for resuscitation in patients with septic shock.
Methods
In this retrospective cohort study, twenty features, including demographic, laboratory, and vital signs data, were utilized as input variables for seven machine learning models. The output variable, defined as the maximum vasopressor dose lasting longer than 30 minutes within the first 2 hours of norepinephrine initiation, was categorized into four or five classes. The models were trained with five-fold cross-validation and evaluated on an independent held-out test set from the Seoul National University Hospital dataset, and externally validated using the Salzburg Intensive Care database.
Results
A total of 670 patients who satisfied the eligibility criteria were included for model development. The proposed model achieved a micro-averaged AUROC and a macro-averaged AUPRC of 0.906 ± 0.002 and 0.373 ± 0.013, respectively, in Class Option 1. In Class Option 2, it scored a micro-averaged AUROC of 0.691 ± 0.012 and a macro-averaged AUPRC of 0.289 ± 0.010. The model also demonstrated robustness in external validation (micro-averaged AUROC of 0.885 ± 0.009 and AUPRC of 0.297 ± 0.012 in Class Option 1, micro-averaged AUROC of 0.674 ± 0.019 and AUPRC of 0.287 ± 0.012 in Class Option 2). Clinicians required an average of 28.5 minutes to reach the target vasopressor dose.
Conclusion
The proposed model may assist clinicians in estimating early vasopressor dose requirements for patients with septic shock. Prospective studies are needed to evaluate its clinical utility.