Fig. 2: Model performance.

a, Receiver-operating characteristic curve for the binary classification task of predicting circulatory failure, comparing the two proposed models with a baseline model. The full model contains 500 features (composed from 112 variables), and the compact model contains 176 features (composed from 16 variables). The baseline model used only variables included in the definition of circulatory failure and is based on a decision tree. b, Precision-recall curve for the circEWS and circEWS-lite alarm systems derived from the full and compact classification model from a. circEWS and circEWS-lite use a 30-min silencing period after every occurring alarm during which no new alarm is triggered. Recall was defined as the fraction of events for which the system correctly raised an alarm from 8 h to 5 min before the event. Precision was defined as the fraction of alarms that are in a window of 8 h prior to a circulatory failure event. c, The fraction of events that correctly trigger an alarm is reported for each 30-min interval during the time period 8 h before circulatory failure occurs. d, Top, the distribution of timing of the first alarm in the 8 h before an event. The mean time from the first alarm to deterioration was 2 h and 32 min. Bottom, the distribution of alarms in 8 h windows that were not immediately followed by an event. In a–c, solid curves were derived from the held-out split, variation estimates were derived from n = 5 independent experiments in the development splits. Prec, precision; rec, recall.