Predicting the Individual Risk of Acute Severe Colitis at Diagnosis

Abstract

Background and aims: Acute severe colitis [ASC] is associated with major morbidity. We aimed to develop and externally validate an index that predicted ASC within 3 years of diagnosis.

Methods: The development cohort included patients aged 16-89 years, diagnosed with ulcerative colitis [UC] in Oxford and followed for 3 years. Primary outcome was hospitalization for ASC, excluding patients admitted within 1 month of diagnosis. Multivariable logistic regression examined the adjusted association of seven risk factors with ASC. Backwards elimination produced a parsimonious model that was simplified to create an easy-to-use index. External validation occurred in separate cohorts from Cambridge, UK, and Uppsala, Sweden.

Results: The development cohort [Oxford] included 34/111 patients who developed ASC within a median 14 months [range 1-29]. The final model applied the sum of 1 point each for extensive disease, C-reactive protein [CRP] > 10mg/l, or haemoglobin < 12g/dl F or < 14g/dl M at diagnosis, to give a score from 0/3 to 3/3. This predicted a 70% risk of developing ASC within 3 years [score 3/3]. Validation cohorts included different proportions with ASC [Cambridge = 25/96; Uppsala = 18/298]. Of those scoring 3/3 at diagnosis, 18/18 [Cambridge] and 12/13 [Uppsala] subsequently developed ASC. Discriminant ability [c-index, where 1.0 = perfect discrimination] was 0.81 [Oxford], 0.95 [Cambridge], 0.97 [Uppsala]. Internal validation using bootstrapping showed good calibration, with similar predicted risk across all cohorts. A nomogram predicted individual risk.

Conclusions: An index applied at diagnosis reliably predicts the risk of ASC within 3 years in different populations. Patients with a score 3/3 at diagnosis may merit early immunomodulator therapy.

Keywords: Biomarkers; acute severe colitis; clinical trials; prediction; ulcerative colitis.

Figure 1. a.

Calibration plot for the multivariable logistic regression in the Development cohort. b. Calibration plot for the final model. These are complex figures for non-specialists that indicate how accurate the model predictions are.^, Outcomes were grouped by fifths of predicted risk. The plots show the agreement between predictions from the model and what was actually observed. The mean predicted risk for each group and the mean observed frequency were calculated, and are plotted as triangles. The lines above and below the triangles are 95% confidence intervals. Perfect predictions should lie on the dashed line [--]. The dotted line […] is a smoothed regression line that broadly lies around the dashed line of perfect fit, indicating good calibration. The dotted line shows this at an individual patient level. At the bottom of the graph, the vertical lines above and below the horizontal line labelled 1 and 0, represent a histogram of predicted risks by whether the patients have ASC [labelled 1: going up] or not [labelled 0: going down].

Figure 2.

The risk of developing acute severe colitis in the 3 years after diagnosis according to the predictive index. The predictive index (whereby one point is assigned at diagnosis for E3 extent of disease, one point for CRP > 10mg/l and 1 point for haemoglobin < 14g/dl [men], or < 12g/d [women], range 0–3) was evaluated in three cohorts: Oxford, Cambridge and Uppsala [Sweden]. The x/N on top of each bar denotes the number of events over the number of individuals with a particular score, so 4/43 means [in the Oxford cohort] that 43 scored 0 of whom four developed ASC.

Figure 3.

Nomogram to predict the risk of developing ASC over the next 3 years from diagnosis of UC for individual patients. For example, consider a male patient with pancolitis at diagnosis [E3: 28 points], a CRP of 23mg/dl [15 points] and haemoglobin 11.4g/dl [53 points], giving a total of 93 points. Check off the total score on the total points line and the individual risk of ASC is ~ 60%. The exact formula for calculating an individual risk of developing ASC over the next 3 years is the logistic regression equation p = 1/[1+exp[-[2.87 + 1.01*[Extent = E3] + 0.02*CRP - 0.35 *Hb]]]. Using the equation, the patient in the example would have a predicted risk of 1/[1+exp[-2.87+1.01+0.02*23-0.35*11.4]] = 0.61 [i.e. risk of 61%].