Marieke van Rosmalen

Reliability of brachial plexus MRI in chronic inflammatory neuropathies 49 3 Figure 3.3 ROC curves per rater 1 – Specificity Sensitivity Receiver operating characteristic (ROC) curve of rater 1 (red line) and rater 2 (blue line) with an area under the curve (AUC) of 0.698 (95% CI 0.539 – 0.858) and 0.804 (95% CI 0.667 – 0.940) respectively. Reference line in green. DISCUSSION This study shows that raters agreed in 26 of 50 (52%) images, indicating poor reliability. Although agreement was better when data were dichotomized (normal versus abnormal), our results indicate that difficulties are related mostly to distinguishing more subtle cases of nerve thickening. Objective criteria for abnormality are needed to avoid false positive and negative results and to optimize the diagnostic value of MRI for inflammatory neuropathies. The poor agreement may have several explanations. We cannot exclude the possibility that the difference in radiological experience between raters underlies the poor reliability. However, the assessors work in the same department and had comparable training in neuroradiology. This may be an indication that interrater variability could even have been higher had we selected radiologists from different hospitals and training backgrounds. Furthermore, the gap in experience represents current clinical practice. Second, assessors of brachial plexus MRI may lack clear reference points, in particular when abnormalities are two-sided, which may have caused best-guessing particularly in cases from category 2. Third, three categories in the scoring model may have been one category too many. Analysis of dichotomized data led to a slightly higher kappa, but still indicated a poor level of