A generalized linear classification model with a smooth link function and predictors obtained from quantile spline fits to high-dimensional data

Abstract

A classification model is developed for grading continuous recordings of the kind produced automatically by an analytical instrument. The explanatory variables of the model are features of the recordings extracted via (quantile) smoothing spline fits. The response variable is a grade and, in the training dataset, is an experts grade for a variety of recordings. A smooth inverse link function, allowing for nonlinearity in the experts grading scale, is incorporated in the model. The procedure is presented by way of an example that involves grading the quality of hypodermic syringes from friction profiles, obtained via tests conducted using an Instron™ machine. The linear predictor of the model describes grades, assigned by the expert to a training set of syringes, in terms of extracted features from associated friction profiles. A wide range of syringe quality is included in the training set. This allows application of the classification model to a large class of quality improvement studies. Using leave-one-out cross-validation, the prediction accuracy of the model is found to be about the same as that of the expert.