Simultaneous Prediction of New Morbidity, Mortality, and Survival Without New Morbidity From Pediatric Intensive Care: A New Paradigm for Outcomes Assessment

Abstract

Objectives: Assessments of care including quality assessments adjusted for physiological status should include the development of new morbidities as well as mortalities. We hypothesized that morbidity, like mortality, is associated with physiological dysfunction and could be predicted simultaneously with mortality.

Design: Prospective cohort study from December 4, 2011, to April 7, 2013.

Setting: General and cardiac/cardiovascular PICUs at seven sites.

Patients: Randomly selected PICU patients from their first PICU admission.

Interventions: None.

Measurements and main results: Among 10,078 admissions, the unadjusted morbidity rates (measured with the Functional Status Scale and defined as an increase of ≥ 3 from preillness to hospital discharge) were 4.6% (site range, 2.6-7.7%) and unadjusted mortality rates were 2.7% (site range, 1.3-5.0%). Morbidity and mortality were significantly (p < 0.001) associated with physiological instability (measured with the Pediatric Risk of Mortality III score) in dichotomous (survival and death) and trichotomous (survival without new morbidity, survival with new morbidity, and death) models without covariate adjustments. Morbidity risk increased with increasing Pediatric Risk of Mortality III scores and then decreased at the highest Pediatric Risk of Mortality III values as potential morbidities became mortalities. The trichotomous model with covariate adjustments included age, admission source, diagnostic factors, baseline Functional Status Scale, and the Pediatric Risk of Mortality III score. The three-level goodness-of-fit test indicated satisfactory performance for the derivation and validation sets (p > 0.20). Predictive ability assessed with the volume under the surface was 0.50 ± 0.019 (derivation) and 0.50 ± 0.034 (validation) (vs chance performance = 0.17). Site-level standardized morbidity ratios were more variable than standardized mortality ratios.

Conclusions: New morbidities were associated with physiological status and can be modeled simultaneously with mortality. Trichotomous outcome models including both morbidity and mortality based on physiological status are suitable for research studies and quality and other outcome assessments. This approach may be applicable to other assessments presently based only on mortality.

Figure 1

Morbidity (Figure 1A) and Mortality (Figure 1B) for Dichotomous (dashed lines) and Trichotomous (solid lines) Relationships. The graph illustrates the relationships in the development set (n = 7650). There were 113 patients with PRISM III scores > 20. There was a significant association of morbidity and mortality with physiological instability measured with the PRISM III score (p<0.001 for likelihood ratio test of significance) for both outcomes.

Figure 1

Morbidity (Figure 1A) and Mortality (Figure 1B) for Dichotomous (dashed lines) and Trichotomous (solid lines) Relationships. The graph illustrates the relationships in the development set (n = 7650). There were 113 patients with PRISM III scores > 20. There was a significant association of morbidity and mortality with physiological instability measured with the PRISM III score (p<0.001 for likelihood ratio test of significance) for both outcomes.

Figure 2

Volume Under the Surface (VUS) for the Trichotomous Predictor. Data are shown for the Derivation Set. The three edges of this surface are the two two-dimensional ROC curves for the prediction of each pair of outcomes. In general, vertical “slices” of this surface in any dimension may be viewed as conditional ROC curves showing the ability of the model to predict a pair of outcomes conditional on correctly predicting the third outcome in a given proportion of subjects. Estimated VUS (proportion of the “cube” that is under the prediction surface) was 0.50 ± 0.019 for the derivation set and 0.50 ± 0.034 for the validation set. The shaded triangular space indicates chance performance analogous to the diagonal line in a two-dimensional ROC curve which in three dimensions is a VUS of 0.167. Average dichotomized c-index (chance performance=0.50) was 0.84 ± 0.009 and 0.85 ± 0.016 for the derivation and validation sets.

Figure 4

Standardized Morbidity (Clear Bars) and Mortality (Solid Bars) Ratios with 95% Confidence Intervals For The Individual Sites. Site order does not correspond to Table 2. * = p<0.05 for overall model fit at the indicated site.