Data Driven Tumor Marker Prediction System

In this paper a system for the prediction of tumor marker values based on standard blood is presented. Several neural networks are used to learn from blood examination measurements and predict tumor markers in case these values are missing. In a post processing step the predicted values are evaluated in a fuzzy logic like style against different hypotheses and the best hypothesis is used to optimize the predicted values and its plausibility. These predicted values can then be used as input for a second system to support decision making in cancer diagnosis. A variety of experiments with tumor marker C153 show that we can get a prediction accuracy of more than 90%. Our experiments are based on hundreds of samples of up to 27 different features (blood parameters) per vector. We try to predict distinct values, classes of values and a combination of classes and values for specific marker types.