• 1. Potential of Data Pattern Classification

  - The CLASSIF1 data sieving algorithm permits the development of standardized, instrument and laboratory independent cytometric and other multiparameter classifiers from flow cytometric list mode, flow bead array, mRNA or protein expression chip array (Lymphochip, Affymetrix), clinical chemistry, biomedical or clinical data for predictive medicine by cytomics
  or for diagnostic purposes (literature references: bioinformatics, medical and clinical cytomics)

• 2. CLASSIF1 Algorithm

  - The CLASSIF1 classification algorithm operates as follows:
  - Percentile pairs e.g. 5/95%, 10/90%, 15/95% etc. are calculated for the value distribution of the reference samples of each database column of the learning set (fig.7)
  - The values of each data column are transformed into triple matrix characters (fig.8) by assigning: (-) to values below the lower percentile, (0) to values between the lower and upper percentiles and (+) to values above the upper percentile of the value distribution of the reference sample.
  - A confusion matrix is then established between the known predictive or diagnostic clinical classification of reference and abnormal samples on the ordinate against the same classification categories for the CLASSIF1 triple matrix classifier on the abscissa (fig.9). Correct classification for all samples is indicated by 100% recognition values in the diagonal boxes of the confusion matrix and 0% values in the non diagonal boxes. The initial classification of learning sets is usually significantly away from this ideal result (fig.10A).
  - The iterative CLASSIF1 optimization is directed towards an increase of the sum of the diagonal box values above the initial sum of values when all database columns are classified (fig.10B). Towards this goal, single database columns or variable combinations of database columns are temporarily excluded from the classification process to see whether their absence improves or deteriorates the classification result. Once no significant classification improvement is reached any more, database columns which neither alone nor in combination with other database columns have improved the classification result are definitively excluded, that is only the discriminatory database columns remain as disease classification masks (fig.3) for the normal as well as for the abnormal classification categories.
  - Reclassification of the learning set (fig.2) or the correct prospective classification of unknown test patients (fig.6) shows the validity of the concept.
  - Triple matrix classifiers are inherently standardized onto the reference samples during the classification process (standardized multiparameter data classification (SMDC)). Classifiers can therefore be compared in an instrument and laboratory independent way, in case no differences between the various reference groups are detected by the CLASSIF1 algorithm. This is advantageous for consensus formation e.g. on leukemia, HIV and thrombocyte classifications by immunophenotyping.
  - The performance of triple matrix classifiers depends on intralaboratory precision rather than on accuracy since measurement accuracy cancels out within certain limits through the normalization of the experimental values on the respective mean values of the reference samples in each database column. Reference groups are typically constituted from age and sex matched patients.

• 3. Examples

  - FCS1.0 and 2.0 list mode files or dBase3 database exports from database (e.g. Access) or table calculation programs (e.g. Excel) are classified by the CLASSIF1 algorithm in various clinical situations:

  myocardial infarction risk assessment (thrombocyte activation antigens)

  overtraining risk in competition cyclists (lymphocyte antigens)

  survival prognosis in melanoma patients (clinical data)

• 4. Database Download

  Comparative multiparameter data classification efforts are significantly hampered by the lack of suitable databases. The downloadable CLASSIF1 database (PHLUNG1.XLS) of flow cytometric intracellular pH and esterase activity measurements in vital cell preparations from malignant and adjacent healthy cells of lung cancer patients may serve as an example for comparative classification efforts. Three parameter ADB/PI/electronic cell volume list mode files were exhaustively analysed as described by: Liewald et al. Cytometry 11:341-348(1990).
  

• 5. Conclusions

  - the CLASSIF1 algorithm provides access to Predictive Medicine ( >95% correct disease course prediction in individual patients) as well as to standardized diagnostic classifications.
  - the CLASSIF1 approach facilitates the elaboration of interlaboratory consensus classifiers in complex multiparameter data sieving or data mining analysis.
  - as a practical consequence, diseases can be classified at institutions where no sufficient learning sets can be generated in reasonable times or where costly investigations are necessary to establish appropriate learning sets.
  - furthermore the molecular and biochemical properties of many body cell systems during disease can be compared by standardized classification e.g. blood leukocytes versus tissue or effusion leukocytes.

Off-line Internet, a timesaver !