Urinary protein markers in lupus nephritis: the need for concurrent calibration and discrimination statistics in predictive models

To the Editor: Researchers increasingly investigate artificial neural networks as a tool for individualized prediction in clinical medicine. A recent article by Oates et al.1 in Kidney International, concluded that an artificial neural network could identify significant urinary protein spots for predicting class and chronicity in patients with lupus nephritis. However, only the discrimination statistic was used. Measurement of the calibration statistic, a key step to evaluate the goodness-of-fit reliability for neural networks was ignored. Without both statistics, readers cannot recognize whether the model is truly applicable.