7.P.59 THE USE OF ARTIFICIAL INTELIGENCI FOR PREDICTING THE BIOLOGICAL BEHAVIOR OF BONE TUMORS

Abstract

Introduction: The prediction of clinical and biological behavior of bone tumors plays an important role in medical tasks such as diagnosis and treatment planning. Different prognostic factors for bone tumors outcome appear to be significant predictors for making definitive diagnosis. It is well-know that different clinical, radiological and histological characteristics are included in diagnostic process. The most important task for pathologist is to determinate biological behavior. Errors in diagnosis lead to wrong therapy and treatment.

It was reason to determinate scores for tumor diagnostics. Score is usually determinate using classic statistical methods such multivariate logistic regression (MVLR), but new computer tehniks, and models of artificial intelligence take a place in modern scoring systems. Recently, classifications tree analysis (CTA) and artificial neural network (ANN) models have become popular in decision-making and outcome prediction of clinical medicine, especially in oncology.

This study compared the levels of accuracy of MVLR, CTA and ANN model for the prediction of bone tumor’s biological behavior.

Material and method: Data from patient who had diagnosed bone tumors in Institute of pathology, School of Medicine in the period of 10 years (1995–2004) were used for analysis purposed in the study. In the analyzed date –base were 3689 biopsies with these criteria. About 24% (882 biopsies) were excluded because of missing data about radiological presentation. Consequently, data from 2807 biopsies were used for the analyses

Clinical, radiological, histological characteristics, summary 166 variables were analyzed and used to compare the levels of accuracy for the three methods of scoring.

All data were inserting in Spider 2.0 enterprise date-base who assisted MSSQL server 2000.

For MVLR and CTA we used SPSS 15.0 program with incorporate CTA. There are methods of multivariate analysis that allow for study of simultaneous influence of a series of independed variable on the one depended variable (biological behavior of bone tumors). The ANN model used in this study were feed-forward networks, witch were trained with a back propagation algorithm (NNSYSID-Neural Network Based System Identification Toolbox) situated in the Matlab area.

We compared three models across theirs overall percentages. The best model was one with highest overall percentage.

Results: From all analyzed cases 1590 (56, 6%) were males and 1217 (43, 4%) were females patients with Middle Ages 34, 1 (aged from 0–94 years). Malignant bone tumors (prime and metastatic lesions) were 1339 (47,7%) and benign 1468 (52,3%).

From all (166) characteristics 11 were selected on the bases of a definitive analysis and included into scoring system. From clinical characteristics just age of patient and clinical diagnosis “cyst” were included. Next radiological presentations: Pure osteolysis, osteolysis with cortical destruction, osteolysis with soft tissue mass, mixed lytic and sclerotic lesion was statistically significant for scoring model. Histological presents of fibroblasts, giant cells with hamosiderin pigment in stromal cells and atypical stromal cells, and hondroid stromal production were important for classification. Localization in finger’s bone was included in definitive score too.

Three performed scoring models showed wary high overall percentages in prediction biological behavior of bone tumors: MVLR 93, 77%, CTA 88, 2% and ANN 91, 5%. The most informative variable, rang 1 in both models of artificial intelligence was radiological criterion. For CTA it was radiological presents of lytic lesion with soft tissue mass and for ANN was combined lytic and sclerotic presentation.

Conclusions: All three scoring models are very useful in prediction bone’s tumor behavior, most of them each ones had priority versus others. The most successive (overall percentage 93, 77%) was MVLR. ANN had high sensitivity (overall percentage 93, 77%) and gave ranges of variables included in score. CTA algorithm had the least overall percentage but it is very simple and figurative for interpretation.