NEW DIAGNOSTIC CRITERIA FOR POLYCYTHEMIA RUBRA VERA: ARTIFICIAL NEURAL NETWORK APPROACH

Submitted to ISCA 15th INTERNATIONAL CONFERENCE ON COMPUTERS AND THEIR APPLICATIONS (CATA-2000) March 29-31, 2000, New Orleans, Louisiana, USA

NEW DIAGNOSTIC CRITERIA FOR POLYCYTHEMIA RUBRA VERA: ARTIFICIAL NEURAL NETWORK APPROACH

Mehmed Kantardzic*, Hazem Hamdan*, Benjamin Djulbegovic**,Adel S. Elmaghraby* *Multimedia Research Laboratory, CECS Department, Speed Scientific School, University of Louisville, Louisville, KY **Division of Blood and Bone Marrow Transplant, H.Lee Moffitt Cancer Center & Research Institute, University of South Florida, Tampa, FL ABSTRACT An ANN-based approach in optimal selection of input parameters for Polycythemia Vera (PV) diagnostics and classification based on this new input set is presented in this paper. Results showed that a trained artificial neural network, known as a robust modeling system, gives the solution for classification of “gray” PV diagnostic zones that occur in practice, and therefore it obtain high quality decisions not only for PV diagnostics, but also for selection of lab data as PV diagnostic symptoms in clinical practice. Ten lab and other clinical findings on 431 PV patients from original PVSG cohort and records on 91 patients with other myeloproliferative or secondary polycythemia were included in this study. Significant differences (p= 36 ml/kg BW Female>=32 ml/kg BW A2: Arterial saturation (OSAT) O2>92%

A3: Splenomegaly

(SPLEEN)

DIAGNOSIS OF POLYCYTHEMIA VERA:

CATEGORY B Lab Tests B1: Thrombocytosis (WBC) >400,000/mm3 B2: Leukocytosis>12,000/mm3 (PLAT) (no fever or infection) B3: Leukocyte alkaline phosphatase (LAP) Score>100 B4a: Serum vitamin B12 >900pg/ml or (B12) B4b: Serum UB12BC >2,200 pg/ml (UBBC)

A1 + A2 + A3 or A1 + A2 + any two from category B

Table 1. Polycythemia Vera study group criteria (PVSG) for diagnosis of Polycythemia Vera

In this paper we present the results of a methodology for PV diagnosis which is based on artificial neural network technology. Our initial hypothesis supposed that a trained ANN, known as a robust modeling system, gives the solution to “gra diagnostic zones, which occur in practice, and therefore obtain high quality of classification for patients with and without PV disease using lab and other clinical findings [8]. Based on trained ANN results we have additional possibilities to analyze the input parameters and their influence on diagnostic process [9]. Optimal selection of inputs is one of the most important results presented in this paper. 2.

ARTIFICIAL NEURAL NETWORKS METHODOLOGY

A process of feature extraction represents selection of input parameters for Polycythemia Vera diagnostics. At the beginning of our research we selected ten parameters as input dimensions for ANN training. Selection is based on experience with common used PVSG criteria and recommendations in previous studies [5]. The parameters are coded in further tables with: sex, REDMAS, OSAT, SPLEEN, HCT, WBC, PLAT, LAP, B12, and UBBC, and they are similar to the set of lab findings for PVSG criteria. Additional parameter, which was not included into PSVG criteria, but it was used in Polycythemia Vera diagnostics in everyday clinical practice, is HCT.

Submitted to ISCA 15th INTERNATIONAL CONFERENCE ON COMPUTERS AND THEIR APPLICATIONS (CATA-2000) March 29-31, 2000, New Orleans, Louisiana, USA

To develop our ANN-based methodology we used the data about patients with Polycythemia Vera and with several similar diseases. The data were collected during last several years, and the total of 522 patient records was included into analysis [5,7]. Because of large amount of missing data for some patients, several preprocessing phases are performed and initial set was transformed. Out of initially collected 522 patient records we continue our preprocessing with 424 records, which have three and less than three missing parameter values, an they were transformed into 1654 “virtual” records. Details about applied ANN methodology and the process of “virtual” record generation are given in [8]. 98 records with more than three missing data were eliminated from further analysis as not enough informative patterns. Implementation of artificial neural network was based on Stuttgart Neural Network System (SNNS V#4.1), software package available on HP graphical stations under UNIX operating system. The training process were carried out with a standard feedforward / backpropagation neural network with one hidden layer in architecture. The commonly used logistic function was selected as the transfer function, and a learning factor was set to 0.2 [1,6]. The network was trained with 828 input-output patterns, and tests were performed with 826 additional “virtual” records. Testing of classification performances for a trained ANN was realized with a set of additional 826 “virtual” patient records not used in a training process. We repeated a testing process with different ANN architectures with different number of hidden nodes and different output threshold values. Two types of errors (false positive and false negative) were analyzed as a result of an ANN testing process [2]. We selected as an optimal solution a trained network with 4 hidden nodes and output threshold value of 0.6. The total number of classification errors in this case was ER=14/826 = 0.016 (or < 2%). Using algorithm described in [8]we transformed 826 tested “virtual” records into 206 original records with missing input data. Output classification values are also automatically computed. Final results of ANN testing using original records with missing parameter values are given in Table 2. Expected Values

Predicted Values

Table 2.

1

0

Total

1

172

2

174

0

1

28

29

not decided

0

3

3

Total

173

33

Results of ANN testing with original data set

Submitted to ISCA 15th INTERNATIONAL CONFERENCE ON COMPUTERS AND THEIR APPLICATIONS (CATA-2000) March 29-31, 2000, New Orleans, Louisiana, USA

Correct classification of patient records in 97.1% cases supports our initial assumption that ANN technology because of its robustness gives very competitive results in a diagnostic process of Polycythemia Vera comparing with other decision making techniques including PVSG criteria. Significant differences (p