Abstract
Dear editor,
We read the paper by Parmar et al. (1) with concern. There are several issues which either appear incorrect or very unlikely.
Discrepancies between the numbers of cases in the text and tables. Text of results states there were 752 incident stroke events arising from 80 308 pyrs giving a crude incidence rate of 9·4 per 1000 pyrs. Table 2, however, states there were 410 events in ARCOS; 45 events in Russia and 675 in the Rotterdam dataset. The total from the table is therefore 410 + 45 + 674 = 1130 not 752 as reported in the text. This gives a crude incidence rate of 14·1 per 1000 pyrs. Either way, the crude incidence rates indicate an extremely high-risk population for a primary prevention study Incorrect calculation of QStroke scores – variables used. We are concerned that the QStroke score has not been calculated correctly. The information in table 2 overlooks three variables which are required for calculation of the score (valvular heart disease, cholesterol/HDL ratio, and deprivation). The table also includes a variable which is not part of the QStroke score (LVH) Incorrect application of QStroke scores to incorrect age ranges: figure 1 looks decidedly odd for QStroke as it shows that stroke risk hardly changes with age in men. The QStroke score only works for patients aged 25–84 years (2), and it is incorrect to apply it to patients aged 85–95 years as indicated in figure 1. The QStroke score has fractional polynomial values for age and age interactions so this could lead to strange values outside the modeled age range. Discrimination. The text of the results (1) initially comments that QStroke outperformed both Framingham and the Riskometer based on all the metrics on each individual dataset based on the information in table 3. The authors then comment on the metrics for the combined dataset. The ROC value for five-year QStroke in men was 80·6 for Russian cohort and 66·1 for Rotterdam. It is therefore extremely unlikely that the ROC value for QStroke in men in the combined dataset dropped to only 59·7 when the majority of the person years of the combined dataset is based on the Rotterdam dataset. If QStroke performed better in each cohort (as they state), then how could it perform so much worse when combined? We think this is most likely to be an error. It is also unclear how the QStroke score could be legitimately applied to the ARCOS cohort since that only consisted of cases of stroke and only had one-year of follow-up.
We would be grateful if you could review this article closely perhaps with a statistician. We would also like to request that original data so that we can verify the results.