External validation of a machine learning model to predict hemodynamic instability in intensive care unit

Figures

• 

  Fig. 1. Flow diagram for inclusion and exclusion criteria of patients

• 

  Fig. 2. Model and threshold performance plots. a Time-varying results of HSI model at different prediction times before hemodynamic interventions, compared to Shock Index and systolic blood pressure (systolic BP); b median comparison between the US cohort and TPEVGH cohort (not all the features are listed due to complications of whole figure); c HSI recall–precision curve of TPEVGH cohort; d HSI recall–specificity curve of TPEVGH cohort

• 

  Fig. 3. Time-varying true alarms and leading time plots. a The fraction of events that correctly trigger an alarm is reported per hour in 24 h before any hemodynamic intervention occurs. b The distribution of timing of the first alarm in the 24 h before an event. 95% unstable patients can be identified over 5 h in advance to interventions

• 

  Fig. 4. Model performance in different subgroup cohorts. a Model performance in different admission source subgroups of TPEVGH cohort. * means outliers, identified by 1.5*IQR; b model performance in different subgroups of TPEVGH cohort on gender, admission type, surgical status, and death. * means outliers, identified by 1.5*IQR; c HSI model AUROC performance by year of TPEVGH cohort