Classification of postoperative cardiac patients by means of simple neural networks

The aim of this work is to compare the performance of simple neural networks with that of statistical Gaussian classifiers in the discrimination of normal- and high-risk post-operative cardiac patients. The considered data include 13 cardio-respiratory quantities taken from 158 patients. Feature extraction techniques were applied in order to extract the 3 variables out of 13 more suitable for classification. The results indicate that there are combinations of three original variables which allow better discrimination between normal and high-risk patients by neural networks than by a Gaussian classifier. For instance, using the cardiac index, the left ventricle stroke work index and oxygen pressure in mixed venous blood, a linear perceptron can discriminate patients with an error as low as 6.9%. Application of the Karhunen-Loeve transform allows the classification error to be reduced to 5.1% with both neural and statistical classifiers if a non-linear discriminant function is used.