Electronic Letters to:
|
|
eLetters: Electronic letters published:
-
![[Read eLetter]](/icons/misc/sectionDOWN.gif)
QRISK - Limitations in Validation?
- Marie Therese Cooney, Alexandra Dudina, Prof Ian Graham, Professor of Preventive Cardiology (8 May 2008)
|
|
|||
|
Marie Therese Cooney, Research Fellow in Cardiolgy Adelaide Meath Hospital Dublin, Alexandra Dudina, Prof Ian Graham, Professor of Preventive Cardiology
Send letter to journal:
Marie-Therese.Cooney{at}amnch.ie Marie Therese Cooney, et al.
|
Dear Editor, We have read with interest the recent article by Hippisley Cox et al on the validation of the new risk estimation function, QRISK. [1] QRISK demonstrated improved discrimination and calibration compared to Framingham in the independent British database, THIN. Previously, we wrote suggesting some methodological limitations in the derivation of the function. [2] Our main concerns were firstly, the use of data in which 70% had missing values for total cholesterol/HDL cholesterol ratio which were, su8bsequenrtly imputed and secondly, the use of data from persons on statin therapy. [3] We believed that the factors were contributing to the trivial and insignificant hazard ratio associated with TC/HDL ratio in the original function (0.001 per unit increase in ratio). Hippisley-Cox et al subsequently published a very satisfactory reply in which these concerns were confirmed. [4] A new version of the function excluding those on statin therapy and using an improved multiple imputation method was derived in which the TC/HDL ratio was a significant risk factor with a hazard ratio of 1.20 in men and 1.17 in women for CVD events. This was the version used in the recent validation exercise. In a further version of the function in which all those with missing data were excluded, the TC/HDL ratio assumed greater importance again, with hazard ratios of 1.25 in men and 1.20 in women. We suggest two possible limitations of the validation exercise. Firstly, again a large proportion of the dataset had missing data on lipid measurements (29%). In the case of missing data in the validation cohort mean values from the derivation cohort, QRESEARCH were substituted. The authors have not reported whether there were differences in outcomes in those with missing data. However, substantial differences can be assumed, given that in the previous cohort the mortality in those with missing data was substantially higher than that of individuals with complete data. (10.9% versus 4.9% respectively) Therefore, it is not reasonable to assume that this group would have the same risk factor levels as those with complete data in the QRESEARCH database. We suggest that validation of QRISK in a cohort with more complete data should be undertaken prior to widespread acceptance of the function. Secondly, the comparison to the performance of the Framingham function has been done using the function which was derived in 1991, using North American data with baselines between in the 1950s and 1970s. [5] Obviously, important secular trends have occurred during the intervening years. Therefore, the improvement in calibration of the QRISK function compared to Framingham may only reflect the use of more recent and local data, as opposed to an improvement related to the methods or variables used in QRISK. We suggest that a comparison to a version of Framingham or SCORE re-calibrated to a modern British population would be more appropriate. Additionally, while improve in the AUROC is a reasonable indication of superior discrimination of the model, the percentage of individuals correctly reclassified to a different risk category using the new function is particularly clinically relevant, since treatment decisions are often made based on risk categorisation [6] . References 1. Hippisley-Cox J, Coupland C, Vinogradova Y, Robson J, Brindle P 2. Cooney MT, Dudina AL, Graham IM 3. Hippisley-Cox J, Coupland C, Vinogradova Y, Robson J, May M,
Brindle P 4. Hippisley-Cox J, Coupland C, Vinogradova Y, May M, Brindle P 5. Anderson KM, Odell PM, Wilson PW, Kannel WB 6. Pencina MJ, D' Agostino RB S, D' Agostino RB J, Vasan RS |
|||
