Dear Editor,<\/p>\n
We were interested to read the article \u201cPrediction of survival by texture-based automated quantitative assessment of regional disease patterns on CT in idiopathic pulmonary fibrosis\u201d authored by Lee SM and colleagues published in European Radiology recently [1]. The authors investigated whether the baseline extent and 1-year change in regional disease patterns on CT can predict survival of patients with idiopathic pulmonary fibrosis (IPF). The results were interesting. However\uff0csome methodologic issues should be taken into account.
\nFirstly, the authors selected variables for multivariate analysis when they had P values less than 0.05 in the univariate analysis, which is questionable. The variables in univariate analysis with p-values less than 0.05 which have great effect on result and p-values less than 0.2 may have a relatively small effect on result. If we selected only those with p-value<0.05 in multivariable analysis, the effect of such predictors will be overestimated and in following the prediction power will be decreased. So, P values less than 0.2 in the univariate analysis should be used as input variables for multivariate analyses. Otherwise, the testimation bias in the results will be induced [2].
\nSecondly, the Cox proportional regression model was applied in this study to estimate hazard ratios for survival in patients with IPF. However, the proportional hazard assumption, the main and vital assumption for this model, has not been verified. The valid outcomes cannot be made through the Cox proportional regression model when the proportional hazard assumption is violated [3]. The proportional hazard assumption should be tested by some appropriate methods before Cox proportional regression [3], otherwise, misleading results may be resulted.<\/p>\n
Finally, the authors mentioned that texture-based, automated CT quantification of fibrosis can be used as an independent predictor of survival in IPF patients. However, such conclusion can be misleading due optimism. The prediction models need to be validated internally or externally through using appropriate methods such as cross-validation and bootstrapping; otherwise, their results would be optimistic interpretation [4].<\/p>\n