Development of a Diagnostic Support System for Hematological Disorders Using a Self-Organizing Map.

Though crucial in determining treatment and prognosis, rapid and precise diagnosis of hematological disorders is often difficult. The fact that there is not always a hematologist in every medical institution is another reason why the development of a diagnostic system for hematological disorders is desirable. We used a self-organizing map (SOM) to establish a diagnostic support system for hematological disorders and investigated its usefulness. Data on peripheral blood (CBC, Ret, Diff), marrow differential cell count, immunophenotyping, and cytogenetic data from 402 patients diagnosed with hematological disorders between 1995 and 2003 were submitted into this system. These included 127 cases of acute myeloid leukemia (AML) 50 cases of acute lymphocytic leukemia (ALL), 17 cases of biphenotypic acute leukemia (BAL), one case of large-granular lymphocytic leukemia (LGLL), 55 cases of mature-B cell neoplasms (mature-B), 9 cases of mature-T cell neoplasms (mature-T), 43 cases of multiple myeloma (MM), and 100 other cases. The other cases consisted of 49 cases of myelodysplastic syndrome (MDS), 8 cases of myeloproliferative disorder (MPD), 26 cases of chronic myeloid leukemia (CML), 13 cases of idiopathic thrombocytopenia purpura (ITP), and 4 cases of anemia. The patients were classified into eight main groups (AML, B-ALL + BAL, T-ALL, LGLL, mature-B, mature-T, MM, and others) and a SOM of the main groups drawn. As SOMs express multi-dimensional data in a two-dimensional format, the map used was one with a toroid shape in which top and bottom and left and right are joined together. For subclassification, the multilayer map method was used to allow more detailed divisions. The method of display used was not dots; instead domains were established to include other related diseases in the surrounding area, and the probability of a specific blood disease was displayed, so that the classification made the relationship with the surrounding area visually comprehensible. Through a series of such modifications of SOM Ver.1.0, the handling of diseases, test results, and algorithms was also upgraded, and at present SOM Ver.4.0 is under development. For validation, an evaluation of the usefulness of Ver.4.0 was carried out using data from 65 fresh cases (AML, ALL, mature-B, MM and others). The same diagnosis as the clinical diagnosis was made in 17 of the 18 AML cases, both of the 2 ALL cases, 6 of the 8 mature-B cases, all 4 MM cases, and all 33 other cases by SOM Ver.4.0. The samples in which the initial SOM diagnosis did not agree with the clinical diagnosis were one AML:M6 classified as other, and two mature-B samples classified respectively as MM and other. Next we carried out validation of usefulness in diagnostic support in an actual medical treatment environment. Diagnoses of 10 newly admitted patients were evaluated by comparing diagnoses by residents, general physicians and hematologists with those by SOM Ver.4.0, resulting in predicted diagnoses of hematological disorders similar to the clinical diagnoses. All physicians who participated in evaluating the predicted diagnoses by SOM found the system to be a very useful tool for the diagnostic process and determining further appropriate action. The system thus appeared to constitute a useful diagnostic support tool for all physicians.