Classification models are widely used in ADMET modeling and specificity and sensitivity are two statistical measures often used to gauge such a model's performance.  For many complex endpoints there is a tradeoff between the two, and neither rises to an acceptable level.  When that happens, one may wish to bias the model to selectively reduce the number of false positives or negatives.  One way to do so is to set a minimum specificity or sensitivity.  This results in the desired performance with respect to one criterion, but the complementary statistic suffers.  Alternatively, one can weight the objective function used to train the models so as to favor one statistic over the other.  The presentation will discuss the drawbacks and benefits of the various methods for controlling the trade-off between specificity and sensitivity as implemented in terms of Artificial Neural Network Ensembles.

Controlling specificity and sensitivity in artificial neural network ensemble (ANNE) classification models

243rd National Meeting (2012)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.