Neural network prediction of mortality

A study is made predicting whether a person of age 55+ will survive for ten years, based on the person's answers to eighteen health related questions. The prediction accuracies of a multilayer perceptron are shown to depend greatly on the initial network weights, and overfitting is exhibited as the number of hidden units increase. The best result obtained is 81% on the test set, significantly better than 63% using a classic nearest neighbor classifier. In addition to the intrinsic interest of this medical application, other intriguing features are: 1) some of the questions are not answered by all study subjects, raising the issue of how those answers should be addressed, 2) the question answers contain both symbolic integers and integer measurements of varying magnitudes, raising the question of how to properly weight the different variables.