Neural network approach to incomplete data applied to assessing cardiac health

The project is based on data from the Cleveland Clinic Foundation Clinic, located in Cleveland. In the database, there are 13 variable: age, sex, type of chest pain, resting blood pressure, serum cholesterol, blood sugar levels, results of the resting ECG, maximum heart rate, angina, decrease the value of the ECG ST , slope of the ST segment on the ECG, number of large blood vessels, scintigraphy result. The patient is assigned to one of the two groups: healthy or sick (0 or 1). Using a neural network MLP (Multi Layer Perceptron) with backpropagation learning method, for all 13 parameters almost 95% of correct classification of validation set was achieved. Unfortunately, even a best chosen neural network is not suitable for classification of incomplete data. With the help of genetic algorithm used to select the input group, the most important parameter was found. Maximum heart rate determines the classification of a fairly good result (71.5%). Most of the databases have this parameter. Other easily available parameters were added in order to improve the quality of classification. A choice of four parameters gives the best optimal results for test databases, within the limits of 80% positives, and for one of them even close to 90%. The results demonstrate the possibilities of neural networks to classify vectors of incomplete content.