Analysis of clinical features and construction of prediction models for hemodynamic instability

Yu Haibo; Wu Mingzheng; Dai Shuai; Xia Jian; Jiang Cheng; Zhao Yan

doi:10.3969/j.issn.1002-1949.2024.06.007

Abstract：Objective To investigate the importance of different clinical features in predicting hemodynamic instability,and to construct best prediction model for different application scenarios.Methods Patients admitted to the intensive care unit(ICU)who met the inclusion criteria in the Medical Information Mart for Intensive Care-Ⅳ(MIMIC-Ⅳ)database were retrospectively selected and divided into hemodynamic instable group and hemodynamic stable group according to their clinical interventions.General characteristics,clinical data,laboratory results of all the patients were collected and compared between groups.Machine learning algorithms were used to assess the importance of each feature and construct adult hemodynamic instability(AHI)model,modified AHI model,noninvasive model,invasive model,blood pressure model,and shock index model.Results Model performance was evaluated by using the area under the receiver operating characteristic(AUC)curve,and the AHI model demonstrated the best predictive performance(AUC=0.862).The performance of other models was as follows:the modified AHI model(AUC=0.810),the invasive model(AUC=0.787),the non-invasive model(AUC=0.760),the blood pressure model(AUC=0.720),and the shock index model(AUC=0.716).F1 score showed that the AHI model was best,followed by the modified AHI model,the invasive model and the non-invasive model.The blood pressure model and the shock index model had the worst predictive performance.Conclusions In the case of complete feature information,the AHI model is the best predictive model,and the non-invasive blood pressure is identified as the most useful feature for predicting hemodynamic instability.Under the conditions with limited feature information,the non-invasive model has the advantage over single blood pressure model and shock index model.