Fig. 2.
Calibration curves displaying the relation between observed outcomes and model predictions. Perfect calibration is designated by the 45° line through the origin. Histograms of the risk scores underlie each panel. Calibration curves are truncated to the middle 99% of the data, or to the limits of the displayed axes. (A) Displays the calibration of raw POARisk and AllCodeRisk scores from the logistic model, within the random 20% calibration cohort from 2004 to 2008. Correcting POARisk and AllCodeRisk predictions based on these curves improved calibration but was insufficient to ensure complete calibration in the prospective 2009 data (B). The modified Risk Stratification Index (RSI) tended to overestimate risk in the general California inpatient population. However, recalibration within the 2009 data based on the curves in B yielded favorable calibration for all three models, as displayed in C. POA = present-on-admission.