Paolo Emilio Puddu1, Alessandro Menotti. 1. UOC di Biotecnologie Applicate alle Malattie Cardiovascolari, Department of the Heart and Great Vessels Attilio Reale, University of La Sapienza bAssociazione per la Ricerca Cardiologica, Rome, Italy. paoloemilio.puddu@uniroma1.it
Abstract
AIMS AND METHODS: We investigated 12 763 men enrolled in the Seven Countries Study and 25-year coronary heart disease (CHD) mortality to compare the predictive discrimination of the multilayer perceptron (MLP) neural network versus multiple logistic function based on four standard, continuous risk factors, selected a priori. The patients were grouped according to geographical distribution, which also parallels CHD mortality risk. Logistic model solutions were estimated for each geographic area. Training neural network models were estimated in one high risk (US) and one low risk (Italy) population and each was rerun in each nonindex population. RESULTS: CHD mortality prediction by training MLP neural network or multiple logistic function had similar (0.669-0.699) receiver operating characteristic area under the curve (AUC). The rerun of MLP neural network models derived from the US and Italy yielded comparable AUC similar to the logistic solutions in Northern and Eastern Europe, but higher AUC in two areas [0.633 (logistic) vs. 0.665 or 0.666 (neural network: P<0.05) in Southern Europe and 0.676 (logistic) vs. 0.725 or 0.737 (neural network: P<0.01) in Japan]. CONCLUSION: This is the first investigation performed on epidemiological data to suggest a good performance in predicting long-term CHD mortality, on the basis of few continuous risk factors, of a training neural network model that could be rerun on different populations with satisfactory findings.
AIMS AND METHODS: We investigated 12 763 men enrolled in the Seven Countries Study and 25-year coronary heart disease (CHD) mortality to compare the predictive discrimination of the multilayer perceptron (MLP) neural network versus multiple logistic function based on four standard, continuous risk factors, selected a priori. The patients were grouped according to geographical distribution, which also parallels CHD mortality risk. Logistic model solutions were estimated for each geographic area. Training neural network models were estimated in one high risk (US) and one low risk (Italy) population and each was rerun in each nonindex population. RESULTS: CHD mortality prediction by training MLP neural network or multiple logistic function had similar (0.669-0.699) receiver operating characteristic area under the curve (AUC). The rerun of MLP neural network models derived from the US and Italy yielded comparable AUC similar to the logistic solutions in Northern and Eastern Europe, but higher AUC in two areas [0.633 (logistic) vs. 0.665 or 0.666 (neural network: P<0.05) in Southern Europe and 0.676 (logistic) vs. 0.725 or 0.737 (neural network: P<0.01) in Japan]. CONCLUSION: This is the first investigation performed on epidemiological data to suggest a good performance in predicting long-term CHD mortality, on the basis of few continuous risk factors, of a training neural network model that could be rerun on different populations with satisfactory findings.
Authors: Vera Elizabeth Closs; Patricia Klarmann Ziegelmann; João Henrique Ferreira Flores; Irenio Gomes; Carla Helena Augustin Schwanke Journal: Curr Gerontol Geriatr Res Date: 2017-11-20
Authors: Francesco Macrina; Paolo E Puddu; Alfonso Sciangula; Marco Totaro; Fausto Trigilia; Mauro Cassese; Michele Toscano Journal: J Cardiothorac Surg Date: 2010-05-25 Impact factor: 1.637