Biopsy transcriptome expression profi ling: proper validation is keyPhilip J O'Connell and colleagues (Sept 3, p 983) 1 describe a set of 13 genes they claim has high accuracy for prediction of several chronic allograft damage-related phenotypes in kidney transplants. Because of the large number of genes represented on a microarray chip, rigorous methods must be employed to avoid overfitting. We are concerned that the validation methods in this paper were performed incorrectly, leading to inflated estimates of predictive accuracies.The authors chose 13 genes after extensive selection and fi ltering in the development set, and fi t the models using penalised logistic regression. The authors used two internal cross-validation methods: repeated threefold cross-validation and leaveone-out cross-validation. In applying cross-validation, all steps-including gene selection-must be reapplied within each training set, without using information from the test sets. The authors did this in the leave-oneout cross-validation (area under the curve [AUC] 0•774). However, the threefold cross-validation seems to use the preselected 13-gene set for all models, producing an AUC of 0•889. This resubstitution method is known to generate infl ated accuracy estimates. 2,3 The AUC reported in the Summary and Discussion should not be the unvalidated and clearly overfi t 0•967 actually shown, but 0•774. This value is lower than the 0•81 AUC achieved using histoclinical variables

Patients using 4-FA are at risk for life-threatening health problems, including intracranial haemorrhage. Additional brain imaging should be considered in 4-FA-intoxicated patients, not only in the presence of neurological deficits, but also in the case of severe headache.