Table 3.

Summary of top predictive modelsa,b

ModelbEquationDiscovery accuracyValidation accuracy
A: Top logistic regression model for the combined cohortPredict low risk if expit{13.8712*log10(sgp130Pk3) + 2.4626*log10(IFNγPk3) −1.6559*log10(IL1RAPk3)−75.3502} ≤ 0.3623Sens = 12/14Sens = 2/2
Spec = 31/35Spec = 6/10
B: Top tree model for combined cohortPredict low risk if sgp130pk3 < 218179 or failing that if MCP1Pk3<4636.52 & eotaxinPk3>29.09Sens = 12/14Sens = 2/2
Spec = 34/35Spec = 7/10
C: Top logistic regression model for the pediatric cohortPredict low risk if expit{8.483*log10(IFNγPk3) − 5.599*log10(IL13Pk3) − 16.343*log10(MIP1αPk3) +15.742} ≤ 0.3288Sens = 11/11Sens = 1/2
Spec = 26/27Spec = 8/10
D: Top tree model for pediatric cohortPredict low risk if IL10Pk3 < 11.7457, or failing that if burden <51%Sens = 10/11Sens = 2/2
Spec = 26/27Spec = 9/10
E: Best pediatric classifier using covariates from top pediatric regression modelPredict low risk if IFNγPk3<27.6732 or failing that if MIP1αPk3> = 30.1591 and IFNγPk3 < 94.8873Sens = 9/11Sens = 1/2
Spec = 25/27Spec = 10/10
  • aThe expit function converts the logistic regression score to the modeled probability of being a case. expit(x) = exp(x)/{exp(x) + 1}; Pk = peak; sens = sensitivity; spec = specificity.

  • bThe designation of top regression model considered the overall accuracy in the sensitivity and specificity (prioritizing sensitivity) in the discovery cohort, and preferring models with no fold change factors unless there was an appreciable increase in accuracy with those factors added because they required an extra (baseline) measurement for the cytokines, which did not occur. Supplementary Table S12 provides the full list of models considered. The best tree (classifier) model for the combined and pediatric cohorts stood out as simultaneously maximizing the sensitivity and specificity among the candidate tree models in these respective discovery cohorts. Additionally, in the pediatric cohort we considered the best classifier that did not use burden (model E), a clinical factor that may not be available in some trials.