gms | German Medical Science

Background and problem: Crohn’s disease (CD) and ulcerative colitis (UC) are chronic inflammatory bowel diseases (IBD). The course can be very variable with respect to the type and severity of symptoms, the sequence of relapses and remission, the type of complications and the response to therapy. A disease severity index (DSI, [1]) was published based on clinician expert consensus. We are currently performing a project [2] to examine how this measure reflects patient perceived severity, but agreement and correlation were only fair to moderate. We aim to rebalance the weights of the score with a generalized additive model for location, scale and shape (GAMLSS, [3]).

??????Methods: Survey: Parallel online surveys were conducted of patients and their treating physicians. Physicians provided information for the DSI. This score has 16 items for CD and 13 items for UC. Original item weights range from 1 to 16 based on expert consensus to result in score ranges from 0 to 100 points for both disease types. Patients rated their disease severity on a visual analogue scale spanning 0 to 100. Analyses were to be restricted to pairs where DSI items were complete and both surveys were completed within a maximum time interval of 4 weeks.

Development of the model: We constructed generalized additive regression models to estimate the associations between the patient rating and all items of the DSI per condition (separate models per disease). For the estimation of the coefficients we used a GAM and explored with a GAMLSS in order to account for heteroscedasticity, i.e. changes to the variance and skewness of the response across covariates. Also, we needed to estimate the model with constraints as we required all coefficients to be positive and sum up to the original scale maximum of 100.

Results: Of 341 pairs received, 89 provided data on all items. Measurements of albumin levels were the most commonly missing item, unavailable in the majority of the cases and normal in almost all other cases. This item was therefore removed as insufficiently useful. A total of 76 Crohn’s patients and 71 UC patients remained in the analysis. An ordinary regression of the original DSI resulted in an R2=0.24 with intercept and 0.73 without intercept. Our models using the DSI items had an R2=0.9 for the prediction of the patient rating. However, negative weights were estimated for 6 items in the CD model and 5 items in the UC model. The addition of linear constraints resulted in 10 items with regression coefficients equal to zero for CD and 4 items for UC. The remaining weights had a range from 1 to 32.

Conclusions: We were able to improve prediction of patient ratings of this expert derived clinical score. The constraints led to the removal of many items and, as such, simplification of the tool. A GAMLSS can further improve the precision by also modeling changes in the variance by item.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.