Neural network in hematopoietic malignancies.

BACKGROUND In the last 30 years, the automatization of the hematology laboratory has led to a high increase in the number of full blood counts (FBC) performed daily as well as a very high degree in the quality assessment in terms of accuracy and precision. Moreover, the produced data are so many that their full interpretation needs an expert; often these new parameters are not exploited. The new challenge of the hematology laboratory consists in the translation of numerical data and new parameters into clinical meaningful information also offering, when possible, a pre-diagnostic guideline, maintaining similar high quality level also in these new aspects. METHODS Since the first development of automated cytochemistry for leukocyte differential count, a new efficient pre-microscopic approach to leukemia diagnosis and classification, was made possible at the time of the automated blood cell count. This original method, used by the Bayer hematological H* series and ADVIA120 analyzers, is based on a light assessment of basic cell properties: volume and peroxidase activity (P) in the Perox channel, which is used to differentiate the main leukocyte types according to the different enzyme content and Nuclear density (ND) in the basophil channel, which is used to count basophils and blast cells according the chromatin pattern and to flag the presence of abnormal or immature cells. The simple observation of P and ND two-dimensional cytograms can be used to include any single case in a separate and distinct diagnostic category correlated to FAB and WHO classifications of hematological malignancies. Moreover, on the RBC map, it is possible to obtain some information about the pathological cluster. The simultaneous observation of these three cytograms and the introduction of a simple and quick score provide objective informations on blast lineage, level of myeloid differentiation, chronic versus acute leukemias. Using a simple visual score system, we have reached in a previous study a pre-diagnostic efficiency in 91% of the analyzed samples. ADVIA120 provides cytogram cell distribution at the end of the analytical processes that include the evaluation of 492 signals, the so-called raw data. The working hypothesis is to create a knowledge-based system, an Artificial Neural Network able to directly handle these raw-data for producing classes of diagnostic probability with a high level of efficiency. The Bayer R&D team in Tarrytown has created an ANN software for connecting a set of input data to output through weighted "hidden layers" (i) assessing the 84 ADVIA120 parameter sets to be used, (ii) defining the interim analysis tool for fitting to standard "normal archetypes", (iii) finding a discriminant function normal vs. pathological. We have collected from 22 Italian hematological centers data from peripheral blood of 1000 patients having mainly hematopoietic disorders at diagnosis. RESULTS AND CONCLUSIONS The ANN has been trained with labeled samples. Same analysis has been performed for constructing the "pathological archetypes". Finally, we have tested the ANN model with a small selected set of the collected pathological samples. The preliminary encouraging results show the high capability of this ANN of clustering signals according to the pre-defined normal as well as pathological archetypes. The next big step is to create an application for discriminating different types of anemia, simply using data from patient's peripheral blood.