Cunha Guilherme Bernardino da, Luitgards-Moura José Francisco, Naves Eduardo Lázaro Martins, Andrade Adriano Oliveira, Pereira Adriano Alves, Milagre Selma Terezinha
Curso de Ciência da Computação, Universidade Federal de Roraima, Boa Vista, RR.
Rev Soc Bras Med Trop. 2010 Sep-Oct;43(5):567-70. doi: 10.1590/s0037-86822010000500019.
Malaria is endemic in the Brazilian Amazon region, with different risks for each region. The City of Cantá, State of Roraima, presented one of the largest annual parasite indices in Brazil for the entire study period, with a value always greater than 50. The present study aimed to use an artificial neural network to predict the incidence of malaria in this city in order to assist health coordinators in planning and managing resources.
Data were collected on the website of the Ministry of Health, SIVEP--Malaria between 2003 and 2009. An artificial neural network was structured with three neurons in the input layer, two intermediate layers and an output layer with one neuron. A sigmoid activation function was used. In training, the backpropagation method was used, with a learning rate of 0.05 and momentum of 0.01. The stopping criterion was to reach 20,000 cycles or a target of 0.001. The data from 2003 to 2008 were used for training and validation. The results were compared with those from a logistic regression model.
The results for all periods provided showed that the artificial neural network had a smaller mean square error and absolute error compared with the regression model for the year 2009.
The artificial neural network proved to be adequate for a malaria forecasting system in the city studied, determining smaller predictive values with absolute errors compared to the logistic regression model and the actual values.
疟疾在巴西亚马逊地区呈地方流行性，各地区风险不同。在整个研究期间，罗赖马州坎塔市的年度寄生虫指数是巴西最高的之一，其值始终大于50。本研究旨在使用人工神经网络预测该市的疟疾发病率，以协助卫生协调员进行资源规划和管理。
数据收集于2003年至2009年期间卫生部的SIVEP - 疟疾网站。构建了一个人工神经网络，输入层有三个神经元，两个中间层，输出层有一个神经元。使用了Sigmoid激活函数。在训练中，采用反向传播方法，学习率为0.05，动量为0.01。停止标准是达到20,000个周期或目标值0.001。2003年至2008年的数据用于训练和验证。将结果与逻辑回归模型的结果进行比较。
所有时间段的结果表明，与2009年的回归模型相比，人工神经网络的均方误差和绝对误差更小。
在研究的城市中，人工神经网络被证明适用于疟疾预测系统，与逻辑回归模型和实际值相比，其预测值的绝对误差更小。
