*Identical offset values for signal alignment applied to reloaded and collected data
Kruskal–Wallis nonparametric tests compared LCI values. As differences between the three methods were normally distributed, Bland Altman analysis assessed agreement, and as an additional comparison 95% limits of agreement (LA) were compared to within-session repeatability between the two 3.3.1 collected trials.
All 34 participants contributed complete sets of technically acceptable data: mean (SD; range) age 15 (9;5-47) years, 28 with CF. LCI agreement is summarised in Figure 1. In agreement with Short et al, there was a difference between migrated values vs 3.3.1 collected LCI, albeit smaller in magnitude: 6.50(5.72-9.78) vs 6.79(6.09-10.94),p=0.049; mean (95% LA) difference -0.37 (-1.17-0.66). But unlike in that study, the additional optimisation of methodology resulted in reloaded values which were not significantly different from those directly collected in 3.3.1: 6.86(5.87-10.09),p=0.819; actual difference 0.01(-0.85,0.75) and 0.13(-9.57,9.83)% change. Differences to both migrated and reloaded 3.1.6 data were almost entirely within this study’s within-session repeatability of 3.3.1 collected data [mean 95% LA difference 0.11(-0.83,1.05) and -0.98(-12.0,14.0%), respectively]. We did find evidence of a within-subject software version sequence effect on LCI difference within this dataset, in the same direction of the difference described by Short et al, supporting the randomisation approach taken in this study’s methodology (data not shown).
Collectively the CORCs believe these data together with the three previous datasets3-5 are now sufficient to makes some recommendations to the community. Firstly, that reloaded legacy 3.1.6 data agree strongly with contemporaneously collected 3.3.1 data when study design carefully ensures alignment of all the signal processing factors previously discussed, including ensuring consistent offset values across comparison groups. Secondly, all these studies have demonstrated that, whilst data migration does generate LCI values that differ from prospectively collected 3.3.1 data (due to the lack of O2 drift correction and DDC application), magnitude of difference is relatively small (<0.4 TO, or 5%) and usually within the within-test repeatability. Depending on the reasons for researchers wishing to use longitudinal, legacy data, these differences may be deemed acceptable. Thirdly, variation in findings across these studies highlights that the methodology used to update legacy data can impact results due to varying application of signal processing factors, and care is needed when updating legacy data (whether by migration or reloading), to ensure the method used is fit for purpose, and the process used must be fully transparent and described. The strong agreement with optimised, reloaded data within this study is reassuring for the current GLI normative data project which now specifies that legacy data is reloaded by contributing sites using a detailed protocol to minimise between site differences in approach. As the amount of 3.3.1 collected data contributed to the GLI database increases, the ability to evaluate agreement with legacy reloaded 3.1.6 data will be created and should be evaluated.