یک روش داده کاوی برای بهبود تشخیص پلی سیتمی ورا

This paper presents a data-mining approach to the extraction of new decision rules for Polycythemia Vera (PV) diagnosis, based on a reduced and optimized set of lab parameters. Ten laboratory and other clinical findings (eight parameters from the Polycythemia Vera Study Group (PVSG) criteria+sex and hematocrit (HCT)) on 431 PV patients from the original PVSG cohort, and records on 91 patients with other myeloproliferative disorders that can be easily misdiagnosed with PV, were included in this study. Significant differences were not found in the correctness of diagnostic classification of patients using either a trained artificial neural network (98.1%) or a support vector machine (95%) versus using PVSG diagnostic criteria, which are considered as a ‘gold-standard’ for the diagnosis of PV. Reducing the original parameters of our dataset to only four parameters: HCT, PLAT, SPLEEN and WBC, we still have obtained good classification results. New rules for improved differential diagnosis of PV are specified based on these four parameters. These rules may be used as a complement to the standard PVSG criteria, particularly in the differential diagnosis between PV and other myeloproliferative syndromes.

Polycythemia Vera (PV) is a rare malignant blood disease characterized by an increase in the number of red-blood cells, white-blood cells and platelets in the blood of a patient. PV represents a true neoplasm of the marrow stem cells. The diagnosis of PV is usually made by applying the widely used Polycythemia Vera Study Group (PVSG) diagnostic criteria (Berlin, 1975). Although these criteria were developed for the purpose of entering patients into clinical trials, they have been widely accepted and used in clinical practice (Djulbegovic, Hadley, & Joseph, 1991). These criteria follow the classical form of ‘If…Then’ categorical rules and call for the diagnosis to be established if certain clinical and laboratory data are present (Berlin, 1975, Kassirer, 1989 and Kassirer and Copelman, 1989). The basic rules for the diagnosis of PV are given in Table 1. In fact, the diagnosis of PV is based on ‘all or none’ rule—if the criteria specified in Table 1 are present, then the diagnosis is established. If the criteria are not fulfilled, PV diagnosis is not established and may require further diagnostic testing. Consequently, treatment of PV depends on the fulfillment of this all or none categorical rule.Despite the proven validity of the PVSG criteria (Djulbegovic et al., 1991 and Djulbegovic et al., 1998), empirical data show that practitioners often do not order the tests as stipulated by the PVSG criteria. For example, a recent survey of american hematologists showed that practitioners order the determination of red-blood cell (RBC) volume in 78%, arterial gas saturation in 75%, leukocyte alkaline phosphatase (LAP) score and serum vitamin B12 level in 44% cases only during work-up of patients suspected of having PV (Streiff & Spivak, 1999). One of the suspected reasons for this poor compliance of practitioners with the PVSG diagnostic criteria could be related to the lack of easy access and quick turnaround of the diagnostic tests listed in Table 1. Except for the determination of splenomegaly by the means of physical exam and widely available complete-blood count (CBC) that provides the determination of white-blood cell count (Table 1: diagnostic criterion B1), platelets (diagnostic criterion B2) along with hematocrit (HCT), all other diagnostic criteria are not widely available to the majority of clinicians. In addition, the determination of RBC volume involves radioactive chemicals, and the measurement of arterial gas saturation is an invasive and unpleasant test to perform. Consequently, further facilitation and improvement in our capacity to diagnose PV is needed. In particular, one would like to know what are the diagnostic yield of both CBC and the determination of splenomegaly by palpation, which are two widely available diagnostic tools in the diagnosis of PV. In this paper, we present new rules, based on a reduced set of parameters, which can be used as a complement to the PVSG criteria. We have used standard classification techniques, artificial neural network (ANN) and support vector machine (SVM), for optimal selection of the parameters. Furthermore, we show that the projection of our dataset from the space defined by this reduced set of parameters on the two-dimensional space has two clusters. The new rules are formalized using a decision rules technique. As a matter of fact, our presented methodology is based on common data-mining tools, which can be used for the analysis of any data including that of industrial engineering. The organization of this paper is as follows. Section 2 presents the results of data classification using ANN and SVM techniques. Section 3 provides the steps for the reduction of input parameters, and formalization of new decision rules for PV diagnosis. Section 4 is dedicated to discussion and analysis of our results.

Using common data-mining techniques, such as ANN, SVM and n-dimensional visualization, we have shown that it is possible to make a diagnostic decision about PV with the same level of classification quality, while reducing the number of input parameters only to four: HCT, PLAT, SPLEEN, and WBC. A new set of rules is provided, which is both clinically justified and supported by computer analysis of the available dataset using the state-of-the-art data-mining tools for decision trees and decision rules extraction. These findings could be used to supplement the PVSG criteria, particularly in the differential diagnosis between PV and ET.