Identifying predictors of clinical outcomes using the projection-predictive feature selection—a proof of concept on the example of Crohn’s disease

Abstract

ObjectivesSeveral clinical disease activity indices (DAIs) have been developed to noninvasively assess mucosal healing in pediatric Crohn’s disease (CD). However, their clinical application can be complex. Therefore, we present a new way to identify the most informative biomarkers for mucosal inflammation from current markers in use and, based on this, how to obtain an easy-to-use DAI for clinical practice. A further aim of our proof-of-concept study is to demonstrate how the performance of such a new DAI can be compared to that of existing DAIs.MethodsThe data of two independent study cohorts, with 167 visits from 109 children and adolescents with CD, were evaluated retrospectively. A variable selection based on a Bayesian ordinal regression model was applied to select clinical or standard laboratory parameters as predictors, using an endoscopic outcome. The predictive performance of the resulting model was compared to that of existing pediatric DAIs.ResultsWith our proof-of-concept dataset, the resulting model included C-reactive protein (CRP) and fecal calprotectin (FC) as predictors. In general, our model performed better than the existing DAIs. To show how our Bayesian approach can be applied in practice, we developed a web application for predicting disease activity for a new CD patient or visit.ConclusionsOur work serves as a proof-of-concept, showing that the statistical methods used here can identify biomarkers relevant for the prediction of a clinical outcome. In our case, a small number of biomarkers is sufficient, which, together with the web interface, facilitates the clinical application. However, the retrospective nature of our study, the rather small amount of data, and the lack of an external validation cohort do not allow us to consider our results as the establishment of a novel DAI for pediatric CD. This needs to be done with the help of a prospective study with more data and an external validation cohort in the future.

Fecal calprotectin and platelet count predict histologic disease activity in pediatric ulcerative colitis: results from a projection-predictive feature selection

“…For our statistical analysis, we followed the approach of Wirthgen and Weber et al [25]. Briefly, using the R [29] package brms [30][31][32][33] which is based on Stan [34], we fitted a Bayesian ordinal regression model to our data, and then, using the R package projpred [35][36][37], we performed a projection-predictive feature selection (PPFS) based on this reference model.…”

“…Briefly, using the R [29] package brms [30][31][32][33] which is based on Stan [34], we fitted a Bayesian ordinal regression model to our data, and then, using the R package projpred [35][36][37], we performed a projection-predictive feature selection (PPFS) based on this reference model. The reference model's outcome was the 4-category histologic inflammation as described in the "Assessment of histologic and endoscopic inflammation" section and its predictors were the candidate predictors listed in Supplementary Table S1, but standardized and partially log-transformed as described previously [25]. We emphasize the inclusion of group-level ("random") intercepts for the patient identifiers in the reference model to account for correlation of visits coming from the same patient.…”

“…Accordingly, the aim of this study is to identify a minimal subset of non-invasive parameters that reflect the histologic disease activity in pediatric UC sufficiently enough and to present a feasible method how to predict histologic disease activity in daily clinical practice. To this end, we applied a Bayesian ordinal regression model and a projectionpredictive feature selection, the rationale of which has already been explained in detail previously [25].…”