CLASSIF1 Data Pattern Classification
- The output of the CLASSIF1 learning procedure LRNTHROM comprises the percent (fig.1) and absolute confusion matrices, the reclassification of the learning set (fig.2), the disease classification masks (fig.3), the list of available database parameters (fig.4), the means, SEMs and percentile values of the discriminatory parameters of the disease classification masks (fig.5) and the prospective classification of unknown test patients (fig.6).
- Disease classification masks and patient classification masks are shortly called disease masks and patient masks.
- Normal and risk patients for myocardial infarction as recognised by coronary angiography are correctly identified (fig.1) from the activation antigen pattern on peripheral blood thrombocytes (fig.2).
- The selected parameters in the disease masks (fig.3) derive from all four measurements. The IgG/CD62/CD63/thrombospondin measurements contribute 1/1/1/2 parameters. The discriminatory data columns represent antibody intensity, ratio or surface density parameters. Percent frequency values of antibody negative or antibody positive cells, in contrast, were less discriminatory and not selected from the databases (fig.4). This underlines the importance for the calculation of relative intensities, surface densities and antibody ratios in flow cytometric list mode and histogram analysis.
- The reclassification of the learning set shows that the patient masks (fig.2) do not exhibit systematic differences within each classification category, indicating the stability of the immunophenotype measurements over the collection period of the samples which was about one year.
- The reclassification list shows further that triple matrix classifiers are inherently robust against certain deviations of the patient masks from the ideal disease mask. Correct classification results are obtained although the positional coincidence of: +, - and 0 characters of the patient mask with the selected disease mask is lower than 1.00.
- The unknown test set of patients was a-priori defined as the 1st, 5th, 10th, 15th ... patient of each classification category (normal/infarction risk). The test set (fig.6) was classified very similarly as the learning set (fig.2). This indicates robustness of the CLASSIF1 classifier for the prospective classification of unknown patients.
|© 2003 G.Valet|