Background/Purpose: Evaluation of imaging is important in spondyloarthritis (SpA) research, but loss to follow up often jeopardizes interpretation of the evaluation. The Interpretation may further be challenged by the fact that often different readers have contributed to scores, in multiple read ‘waves’. A common approach is to evaluate patients (pts) with complete follow up (completers analysis), and aggregate scores of individual readers (eg. agreement ≥ 2 out of 3 readers). These approaches are not assumption-free, may cause non-random data loss, and may as such provide spurious estimates and loss of external validity. We investigated if the use of all data in an assumption-free manner (a so called ‘integrated analysis’) affects the precision of estimates for imaging outcomes in pts with axial SpA (axSpA), with completers analysis as reference standard.

Methods: Pts from the DESIR cohort fulfilling the ASAS axSpA criteria were included. Radiographs and MRIs of the SIJ and spine were obtained at baseline (BL), 1, 2 and 5 years. Each film was scored by 2 or 3 readers in 3 ‘reading-waves’ (wave 1: BL only; wave 2: BL, 1, 2 years; wave 3: BL, 2, 5 years). Each outcome was analyzed in two ways: i. according to a ‘combination algorithm’ (‘2 out of 3’ for binary and mean of 3 readers for continuous variables); and ii. per individual reader. The change of each outcome was analyzed by generalized estimating equations (GEE)) with ‘time’ as explanatory variable. Three analytical approaches were pursued: i) ‘integrated-analysis’ (including all pts with ≥ 1 score from ≥ 1 reader from all waves); ii) completers-only analysis (including only pts with complete 5-year follow-up, using scores from individual readers from wave 3); iii) aggregated completers analysis using a combination algorithm (the same as ii but using combined scores).

Results: In total, 413 pts were included (mean (SD) symptom duration: 1.6 (0.9) years) and 366 completed the 5-year follow up. An analysis with all data from different readers and ‘waves’ (‘integrated analysis’) was more inclusive, but did not result in a meaningful loss of precision (width of 95%CIs) of the change-estimates as compared to both completers analyses (table). In fact, for low-incident outcomes (e.g. the formation of new syndesmophytes), an increased incidence was ‘captured’ by the ‘integrated analysis’ but not by the completers analysis with combined scores (% change/year (95% CI): 0.67 (0.34; 1.01) vs 0.39 (-0.10; 0.90), respectively). The same results were seen using continuous outcomes.

Conclusion: An efficient and entirely assumption-free usage of all data from different readers and ‘read-waves’ does not compromise precision of the estimates of change in imaging parameters, and may yield increased statistical power for detecting changes with low incidence. In addition, integrated analysis may protect against attrition bias and avoid bias by ‘convenient choices’.