Prediction of endoscopic activity in patients with Crohn's disease: systematic review and external validation of published prediction models

Endoscopic healing (EH) is associated with an improved long-term prognosis and is therefore considered a key target in the treatment of Crohn’s disease (CD). Assessment of EH requires ileocolonoscopy, which is a costly and burdensome procedure. A non-invasive index, combining several predictors, to predict EH would simplify and improve management of CD in clinical practice. Published non-invasive models predicting EH often lack external validation. We reviewed the current literature for prediction models for ileocolonic endoscopic activity and subsequently compared their discriminatory abilities using two datasets.

Methods
We systematically searched PubMed, Embase, and the Cochrane libraries until 14 February 2018 for all published diagnostic models based on a combination of at least three predictors, for example, symptoms, serological, or faecal parameters, for ileocolonic endoscopic activity or EH in CD assessed by ileocolonoscopy. We subsequently evaluated the discriminatory value (area under the receiver-operating characteristic curve [AUC]) of the identified models in two separate cohorts, that is, the TAILORIX study1 (346 colonoscopies in 155 patients), and the development dataset of the Utrecht Activity Index (UAI)2 (93 colonoscopies in 82 patients). We corrected for clustering per patient employing the Obuchowski method.

Results
After screening 5303 titles, 21 studies reporting on 27 models with ≥3 predictors were identified. The most commonly used predictors, alongside other predictors in the models, were C-reactive protein (n = 18 [67%]) and faecal calprotectin (n = 13 [48%]). Twelve models were reported in sufficient detail for validation; of these, 8 models could be validated: 6/8 in the TAILORIX and 6/8 in the Utrecht Activity Index dataset. For a threshold of endoscopic activity measured by the CD Endoscopic Index of Severity (CDEIS) ≥3, the AUCs of the published models ranged from 0.55 to 0.85 in the TAILORIX dataset, and from 0.59 to 0.77 in the UAI development dataset (figure). When considering the discriminative ability of continuous values of faecal calprotectin the AUC was 0.82 and 0.79, in the TAILORIX and UAI dataset, respectively, and for CRP: 0.75 and 0.80, respectively.

Conclusions
Based on the discrimnatory ability published prediction models display limited benefit over faecal calprotectin or CRP in prediction of endoscopic activity in CD.