Results

Parts of this chapter have been reproduced with permission from Taylor et al.48 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.

Participants

The final cohort consisted of 194 patients, median age 29, IQR 22–44 years (Figure 2). Ninety-three (48%) of the final cohort were male.

FIGURE 2 Participant flow through trial identifying participants by timing of first event during the follow-up period (> 90 days to 5 years). FU, follow-up; LFU, lost to follow-up. Reproduced with permission from Taylor et al.48 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

All patients and patient events were followed up across the full 5-year period. Because it was possible to include non-contactable patients recruited to the METRIC trial in the retrospective group, it was not possible to define the exact split between the two cohorts (as no patient identifiable data could be sent to the clinical trial unit). Direct consent was obtained from 47 METRIC patients.

According to the modified Beaugerie criteria, 42% (81/194) of participants developed disabling disease between 90 days post diagnosis to 5 years from diagnosis (Table 5 and 6). An additional 25 patients were hospitalised within the first 90 days and although they also fulfilled the modified Beaugerie criteria for disabling disease, after discussion by the TMG, these were not treated as events as these patients were being treated immediately at diagnosis, taking priority over consideration of instigating top-down therapy; including events within 90 days of diagnosis would impact on the clinical utility of any prognostic model.

Only 20% (39/194) developed disabling disease according to the Liège criteria and 6% (12/194) developed B2 or B3 disease on the Montreal criteria (see Table 5).

The event rate using Liège and Montreal definitions of disabling disease was insufficient to perform survival analysis/predictive modelling and therefore all data were analysed using the primary outcome measure (modified Beaugerie criteria) (see Table 6).

Of those with an event that is, development of disabling disease according to the modified Beaugerie criteria, the median survival time prior to the event was 0.82 (IQR 0.42–2.75) years. A total of 81 events occurred in the 5-year follow-up period, 42 (52%) in year 1 and 6 (7%) in year 5 (Table 7). Most Beaugerie-defining events were hospitalisation due to a CD flare or complication (48 events; 59%) or due to use of corticosteroids (21 events; 26%; Table 7).

Clinical predictors of disabling disease

The number of participants with the predefined candidate clinical predictors of disabling disease is shown in Tables 8–10.

Prognostic models

A summary of the various models is shown in Figure 3.

FIGURE 3 Summary of the derivation of prognostic models for developing disabling disease according to the modified Beaugerie criteria. Reproduced with permission from Taylor et al.48 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Model B included the prespecified clinical predictors (excluding blood and stool markers). Models A1, A2 and A3 added the MEGS, sMARIA and LI scores, respectively. Model B1 added the blood and stool markers to the baseline clinical model (model B) (see Figure 3).

Years to events by predictors

Table 11 and Figure 4 show the median survival time to developing disabling disease according to the prespecified clinical predictors. The median MBDD free survival time of all participants was 0.82 (IQR 0.42–2.75) years.

FIGURE 4 Scatter plots of years from diagnosis to developing MBDD, stratified by the prespecified clinical predictors. Markers represent patients who developed MBDD, and orange boxes represent median and IQR. (Figure not reproduced here; see the original article.)

Univariable hazard ratio by clinical predictors

The univariable hazard ratio (HR) for the prespecified clinical predictors and MRI scores is summarised in Table 12. Initial requirement for steroid therapy was associated with a higher risk of an event (development of disabling disease) in both original and imputed data.

The multivariable models (models B, A1, A2, A3 and B1) are shown in Appendix 1, Tables 21–25. The corresponding receiver operating characteristic (ROC) curves are shown in Appendix 2, Figure 8.

FIGURE 5 Kaplan–Meier plots of the percentage of MBDD negative participants in low-risk and high-risk groups over years from diagnosis, stratified by RD. (a) Model B and RD1; (b) Model B and RD2; (c) Model A1 and RD1; (d) Model A1 and RD2; (e) Model A2 and RD1; (f) Model A2 and RD2; (g) Model A3 and RD1; (h) Model A3 and RD2; (i) Model B1 and RD1; (j) Model B1 and RD2. Blue lines represent the low-risk group and red lines represent the high-risk group. Data are n. (Figure not reproduced here; see the original article.)

FIGURE 6 Net benefit of prognostic models to predict the development of MBDD within 5 years of diagnosis. a) RD1 (risk definition group 1-top 40% of participants with the greater risk of MBDD); b) RD2 (risk definition group 2-participants with an absolute risk of MBDD ≥10%). (Figure not reproduced here; see the original article.)

FIGURE 7 Net benefit of the various models according to the ratio of TP to FPs RD2 (patients with an absolute risk of developing disabling disease of ≥ 10%). (Figure not reproduced here; see the original article.)

Modified Beaugerie disabling disease event-free Kaplan–Meier for clinical risk groups

The predictive performance of all the models was assessed using the two a priori defined clinical risk groups and shown in Figure 5. For RD1, the high-risk group included the top 40% of participants with the greatest predicted risk from the model. For RD2, the high-risk group included participants with an absolute risk ≥ 10%.

Tables 13 and 14 show the sensitivity and specificity of each model for identifying patients in each of the two predefined risk groups.

Overall, there was no statistically significant difference in sensitivity or specificity between model B and each of the other models for patient risk group 1 (see Table 13). For patient risk group 2, models A1, A3 and B1 had significantly lower specificity than model B (see Table 14). There was no benefit identified by including the MRE scores (MEGS, sMARIA, LI) or blood or stool parameters, demonstrating that inclusion of these measurements does not improve the prediction of disabling disease.

Model performance in a hypothetical 1000 participants

The performance of the various models in a hypothetical cohort of 1000 patients for RD1 (top 40% of patients with the greatest predicted risk of developing disabling disease) is shown in Table 15. For example, for model B, we would predict 402 to have an event within 5 years of diagnosis, of whom we would correctly predict 217, and falsely predict 185.

The performance of the various models in a hypothetical cohort of 1000 patients for RD2 (patients with an absolute risk of developing disabling disease of ≥ 10%) is shown in Table 15.

In Table 16, we present the participants who started biologic therapy within 180 days of their diagnosis of CD. There was no evidence that early biologics were preventative of MBDD.

Net benefit

An illustration of the net benefit of each model is shown in Figures 6 and 7.

For example, permitting one true positive (TP) to five false positives (FPs) means that to correctly identify one participant developing disabling disease, no more than five participants should be identified incorrectly (i.e. risking unnecessary top-down therapy). For this ratio to be clinically acceptable, the harm of delaying top-down therapy for someone destined to develop disabling disease is five times greater than that of commencing unnecessary top-down therapy. In this example, the ‘weight’ of preventing one patient developing disabling disease is five times more important than administering top-down therapy to a patient who will not develop disabling disease.

A net benefit of 18% means that the model would lead to starting biologics in 180 participants per 1000 at risk, with all biologics truly preventing events (see Figures 6 and 7).

The net benefit of prognostic models for predicting development of MBDD within 5 years of diagnosis, using ratios of TP to FP predictions is provided in Appendix 3, Table 26.

For secondary outcome 2, we had planned to use PCA to determine the optimal combination of individual MRE features (see Appendix 4, Table 27). These features would be used in conjunction with predefined clinical variables (Model B) to predict disabling CD. The MRE features selected (n = 20) were those used in the calculation of the sMARIA, MEGS and LI.

We computed scores for principal components (PCs) and found that the first six PCs accounted for approximately 70% of the total variance. We intended to include these PCs along with the clinical variables in a flexible parametric survival model. However, upon incorporating the PCs, the model failed to converge due to significant collinearity issues with the clinical variables. Consequently, we were unable to explore the predictive ability of individual MRE features using this approach.

Additionally, we performed PCA on a prespecified list of MRE features (n = 11) selected as of clinical interest by the study radiologists, and computed four PCs that accounted for more than 70% of the total variance. We present the scree plot (see Appendix 5, Figure 9) and loading table (see Appendix 8, Table 31). As with the full list of MRE features, the inclusion of these four PCs to the clinical variables led to collinearity problems, hindering the convergence of the model (see Appendix 6, Figures 10 and 11, Tables 28 and 29, Appendix 7, Table 30, Appendix 8, Table 31, Appendix 9, Table 32).