Figure 4.
A multi-panel image depicts the performance of different classifiers and feature importance in distinguishing CVID patients from controls. Panel A shows receiver operating characteristic (ROC) curves for five Ridge regression classifiers distinguishing CVID patients from matched controls. The x-axis represents the false positive rate, and the y-axis represents the true positive rate. The legend indicates different classifiers: PheCodes, Simple, Window, PheCodes plus Simple, and PheCodes plus Window. Panel B is a bar graph showing mean absolute SHAP values for all features retained by the PheCodes plus Window model in the CVID-only analysis, ranked by importance. The x-axis represents the mean SHAP value, and the y-axis lists the features. Orange bars represent windowed episode features, and blue bars represent infection PheCodes. Panel C shows ROC curves for the same five classifiers in the pooled CVID/RTX cohort, with the same axes and legend as Panel A. Panel D is a bar graph similar to Panel B, showing mean absolute SHAP values for features retained by the PheCodes plus Windowed episode features model in the pooled CVID/RTX cohort. The x-axis represents the mean SHAP value, and the y-axis lists the features. Orange bars represent windowed episode features, and blue bars represent infection PheCodes.

Windowed episode features combined with PheCodes outperform simple counting and PheCodes alone in distinguishing CVID and CVID/RTX patients from controls. (A) ROC curves from nested fivefold cross-validation showing out-of-fold performance for five Ridge regression classifiers distinguishing CVID patients from matched controls. (B) Mean absolute SHAP values for all features retained by the PheCodes + Window model after variance thresholding in the CVID-only analysis, ranked by importance. (C) ROC curves for the same five classifiers in the pooled CVID/RTX cohort. (D) Mean absolute SHAP values for features retained by the PheCodes + windowed episode features model in the pooled CVID/RTX cohort. Orange bars, windowed episode features; blue bars, infectious PheCodes. Pairwise AUC comparisons were performed using DeLong’s test with Bonferroni correction. PheCode numbers are shown in parentheses.

or Create an Account

Close Modal
Close Modal