Neural network analysis in predicting 2-year survival in elderly people: a new statistical-mathematical approach

We designed this study to test the usefulness of artificial neural networks (ANN) in assessing 2-year survival in elderly persons, and to understand the net's logical functioning, thus determining the relative importance of the single biological and clinical variables which influence survival. ANN are statistical-mathematical tools able to determine the existence of a correlation between series of data and, once 'trained', to predict output data given input data. Although ANN have been applied in various areas of medical research, they have only very recently been applied in geriatrics (Cacciafesta et al., 2000. Arch. Gerontol. Geriatr. 31 (in press)). We built up an ANN to investigate how 17 clinical variables relating to a sample of 159 elderly people affect survival, and the possibility of predicting 2-year survival or non-survival for each single subject. When tested on a sample of 20 elderly people, the trained network gave the correct answer in 85% of the cases. We then extracted the mathematical function that the net used for calculating the output (survival) for each set of input data (clinical variables). Using this formula, we investigated how some clinical variables influence 2-year survival: we found that a low serum cholesterol level is an unfavourable characteristic in relation to survival. We conclude-despite the fact that the sample studied was relatively small-that ANN are useful in predicting 2-year survival in elderly people. The mathematical function we obtained from the net seems useful in determining the relative importance of single variables related to survival.