Maria Lúcia F Penna1. 1. Escola Nacional de Saúde Pública, Departamento de Endimias Samuel Pessoa, Rio de Janeiro, RJ, Brazil. mlpenna@ensp.fiocruz.br
Abstract
OBJECTIVE: To evaluate recurrent neural networks as a predictive technique for time-series in the health field. METHODS: The study was carried out during a cholera epidemic which took place in 1993 and 1994 in the state of Ceará, northeastern Brazil, and was based on excess deaths having 'poorly defined intestinal infections' as the underlying cause (ICD-9). The monthly number of deaths with due to this cause between 1979 and 1995 in the state of Ceará was obtained from the Ministry of Health's Mortality Information System (SIM). A network comprising two neurons in the input layer, twelve in the hidden layer, one in the output layer, and one in the memory layer was trained by backpropagation using the fist 150 observations, with 0.01 learning rate and 0.9 momentum. Training was ended after 22,000 epochs. We compare the results with those of a negative binomial regression. RESULTS: ANN forecasting was adequate. Excessive mortality (number of deaths above the upper limit of the confidence interval) was detected in December 1993 and October/November 1994. However, negative binomial regression detected excess mortality from March 1992 onwards. CONCLUSIONS: The artificial neural network showed good predictive ability, especially in the initial period, and was able to detect alterations concomitant and a subsequent to the cholera epidemic. However, it was less precise that the binomial regression model, which was more sensitive to abnormal data concomitant with cholera circulation.
OBJECTIVE: To evaluate recurrent neural networks as a predictive technique for time-series in the health field. METHODS: The study was carried out during a cholera epidemic which took place in 1993 and 1994 in the state of Ceará, northeastern Brazil, and was based on excess deaths having 'poorly defined intestinal infections' as the underlying cause (ICD-9). The monthly number of deaths with due to this cause between 1979 and 1995 in the state of Ceará was obtained from the Ministry of Health's Mortality Information System (SIM). A network comprising two neurons in the input layer, twelve in the hidden layer, one in the output layer, and one in the memory layer was trained by backpropagation using the fist 150 observations, with 0.01 learning rate and 0.9 momentum. Training was ended after 22,000 epochs. We compare the results with those of a negative binomial regression. RESULTS: ANN forecasting was adequate. Excessive mortality (number of deaths above the upper limit of the confidence interval) was detected in December 1993 and October/November 1994. However, negative binomial regression detected excess mortality from March 1992 onwards. CONCLUSIONS: The artificial neural network showed good predictive ability, especially in the initial period, and was able to detect alterations concomitant and a subsequent to the cholera epidemic. However, it was less precise that the binomial regression model, which was more sensitive to abnormal data concomitant with cholera circulation.