One-year mortality prognosis in heart failure: a neural network approach based on echocardiographic data
J Ortiz,
CG Ghefter,
CE Silva,
and
RM Sabbatini
Centro de Cardiologia Nao Invasia, Sao Paulo, Brazil.
OBJECTIVES. This study sought to assess the usefulness and accuracy of artificial neural networks in the prognosis of 1-year mortality in patients with heart failure. BACKGROUND. Artificial neural networks is a computational technique used to represent and process information by means of networks of interconnected processing elements, similar to neurons. They have found applications in medical decision support systems, particularly in prognosis. METHODS. Clinical and Doppler-derived echocardiographic data from 95 consecutive patients with diffuse impairment of myocardial contractility were studied. After 1 year, data regarding survival or death were obtained and produced the prognostic variable. The data base was divided randomly into a training data set (47 cases, 8 deaths) and a testing data set (48 cases, 7 deaths). Results of artificial neural network classification were compared with those from linear discriminant analysis, clinical judgment and conventional heuristically based programs. RESULTS. The study group included 57 male (47 survivors) and 38 female patients (33 survivors). Linear discriminant analysis was not efficient for separating survivors from nonsurvivors because the accuracy at the ideal cutoff value was only 67.4%, with a sensitivity of 67.5%, positive predictive value of 27.8% and negative predictive value of 91.5%. In contrast, all artificial neural networks were able to predict outcome with an accuracy of 90%, specificity of 93% and sensitivity of 71.4%, for the best artificial neural network. Both clinical judgment and automatic heuristic methods were also inferior in performance. CONCLUSIONS. The artificial neural network method has proved to be reliable for implementing quantitative prognosis of mortality in patients with heart failure. Additional studies with larger numbers of patients are required to better assess the usefulness of artificial neural networks.
