Prediction of infarction development after endovascular stroke therapy with dual-energy computed tomography: Methodological issues
by Siamak Sabour (email@example.com)Djurdjevic T, Rehwald R, Knoflach M, Matosevic B, Kiechl S, et al. (2017) Prediction of infarction development after endovascular stroke therapy with dual-energy computed tomography. Eur Radiol. 27:907-917.
Dear Editor, I was interested to read the paper by Djurdjevic T and colleagues published in Eur Radiol 2017 Mar. After intraarterial recanalisation (IAR), haemorrhage and blood-brain barrier (BBB) disruption can be distinguished using dual-energy computed tomography (DECT). The aim of the authors was to investigate whether future infarction development can be predicted from DECT . DECT scans of 20 patients showing 45 BBB disrupted areas after IAR were assessed and compared with follow-up examinations. Additionally, Receiver operator characteristic (ROC) analyses using densities from the iodine map (IM) and virtual non-contrast (VNC) were performed .
Based on their results, ROC analyses for the IM series showed an area under the curve (AUC) of 0.99 (cut-off: <9.97 HU; p < 0.05; sensitivity 91.18 %; specificity 100.00 %; accuracy 0.93) for the prediction of future infarctions. They concluded that future infarction development after IAR can be reliably predicted with the IM series. The prediction of haemorrhage and of infarction size is less reliable .
However, this result has nothing to do with prediction. First, for prediction studies, we need data from two different cohorts or at least from one cohort divided into two to first to develop a prediction model and subsequently validate it. Misleading results are generally the main outcome of research that fails to validate its prediction models [2-6]. Moreover, AUC, sensitivity, specificity and accuracy are estimates that are used to evaluate the diagnostic accuracy of a model or a single test compared to a gold standard. Finally, in prediction studies, we must assess the interactions between important variables. Final results can be impacted dramatically when qualitative interactions are present [2-6]. This means that most of the time, without assessing the interaction terms, prediction studies will mainly produce misleading messages.
If you would like to comment on this letter, please send your comments together with full contact data to firstname.lastname@example.org. Replies and comments will be reviewed by the Editorial Office and published on this website