prvmalik
commented
4 years ago Some very quick notes:
The metric is used often for comparing predicted and observed PK parameters (e.g. clearance, AUC_inf). Fold error is defined as [Predicted Value]/[Observed Value]. The percent of model-predicted PK parameters falling within 2-fold of the observed PK parameters is then calculated.
The upper bound of the interval is defined by multiplying by the fold-error and the lower bound of the interval is defined by dividing by the fold error. As examples (and commonly used metrics) 2-fold prediction error is defined as the interval between 0.5 to 2.0-fold error. 1.5-fold error is defined as the interval between 0.667 to 1.5-fold error. 1.33-fold error is defined as the interval between 0.75 to 1.33-fold error and so on.
2-fold error on PK parameters is not acceptable much anymore, except in the world of toxicology. In pharmacology, 1.5-fold error is a standard metric for evaluating predictive utility when extrapolating between species and 1.33-fold error is a standard metric for evaluating predictive utility when extrapolating within species (e.g. from adults to children).
Of interest is applying this method to the actual observed concentrations in the PK profile, instead of just the PK parameters. While we know what kind of accuracy is denoted by finding 80% of predicted clearance values within 1.5-fold error from the observed clearances, we do not understand yet what level of accuracy is denoted by achieving 80% of predicted concentrations falling within 1.5-fold error of the observed values (whether more or less accuracy, who knows yet?)
Presenting this data is most often done in a table where the percent of model predictions are shown at 2-fold error, 1.5-fold error and 1.33 or 1.3-fold error for comparison (obviously these will decrease across these intervals).
Another useful plot is the comparison of predicted vs. observed values on log-log x-y plot where the fold limits can be plotted on either side of the line of equality.
wangwei1619
Incei
tobiasK2001
Hi, all I always see some sentences like 'the predicted result is within 2-fold of oberved data' in literature. And in one paper (DOI: 10.1002/cpt.1013. Epub 2018 Feb 2.), it says 'When evaluating the accuracy and acceptability of predictions, a commonly applied criteria is for values to be within 2-fold of the observed values'. So how to evaluate the performance of a predicition if its result is less than observed data? Is '50% of observed result' applied here? By the way, how is this '2-fold' criteria generated?