We retain the remaining classification nodes; however, we use them for estimating class likelihoods instead of for producing individual classification decisions.
The remaining classification rates obtained in each evaluation are given in Tables 4 through 8 for three to eight combinations, respectively.
The remaining classification methods (Support vector machines, C4.5 decision trees, Logistic regression and Multilayer perceptron) as well as the regression methods (Support vector regression, linear regression) utilized for comparison are available in the Weka package [ 17].
The remaining eight classifications were aggregated into five categories (cropland, forest, grassland, urban land, and water) following the level I classification scheme developed by Anderson et al. (1976).
Australian Refined Diagnosis Related Group Version 5.0 (AR-DRGv5.0) codes were mapped to the remaining detailed classification items by one of the authors (RM).
Diagnosis was established on the basis of at least 3 of the remaining ACR classification criteria, improvement of symptoms within 3 days of corticosteroid therapy and the absence of other condition relevant to the patients' symptoms during a 6-month follow-up [ 42].
The remaining three modified classification systems showed good agreement with classification by the treating physician: thrombocytopenia + bleeding or haemoconcentration, κw value 0.64 (95% CI 0.53 0.74); thrombocytopenia + haemoconcentration, κw value 0.67 (95% CI 0.56 0.78); and haemoconcentration + thrombocytopenia or bleeding, κw value 0.70 (95% CI 0.59 0.81).
Amongst the remaining techniques, the classification-based method provides the best quality.
In our present study cognitive functioning expressed in an IQ explained a substantial part of the remaining variance in competent classifications after accounting for age.
For the remaining 295 treatments the classifications 'Susceptible', 'Potential Low-Level', 'Low-level', 'Intermediate' and 'Resistant' given by the IS for every single drug were mapped to 1,1,0.5,0.5 and 0, respectively.
Effectively this procedure builds a classification model with 90% of the spectra and then tests the remaining 10% to assess classification accuracy.
