engleski

Inductive algorithms in thyroid function diangosis

The aim of this paper is to compare results of three inductive algorithms – Artificial Neural Network, Assistant-algorithm and ILLM (Inductive Learning by Logic Minimisation), applied on thyroid function laboratory diagnostics data. Artificial Neural Network consists of three layers of nodes, applies feedforward information dynamics and backpropagation learning algorithm. Assistant-algorithm is Quinlan’s ID3 algorithm modified by Kononenko and collaborators based on Shannon’s expression for information amount necessary to classify an example. ILLM is a minimization based, two class, propositional, rule generating inductive system. Data base comprised results of the routine assays performed at University Hospital “Sestre milosrdnice”, Zagreb. A sample of 1002 patients was described by 12 laboratory tests and 3 factors relevant to the outcome of diagnosis. According to physical examination, experienced physician decides on diagnosis (Dg) of thyroid function state for each patient: euthyreosis (842 cases), hyperthyreosis (104 cases) or hypothyreosis (56 cases). The diagnosis represents class which should be predicted by laboratory tests results, gender, age and information about possible drug therapy. A sample of 1002 patients was randomized into two sets, training set size of 70% and test set size of 30%. Final results based on training data set are trained Artificial Neural Network, Assistant algorithm decision tree and ILLM production rules. Results are compared by calculating the absolute and relative classification accuracies on the test data set. Absolute accuracies of Neural Network, Assistant-algorithm and ILLM-algorithm predictions were 90.4%, 91.7% and 90.6% respectively, while relative accuracies were 74.7%, 69.1% and 76.4% respectively.

Inductive algorithms in thyroid function diangosis

nije evidentirano