Author:

Keywords:

11 Medical and Health Sciences, 17 Psychology and Cognitive Sciences, Psychiatry, 3202 Clinical sciences

Abstract:

ABSTRACT BODY: Background: In this study we wanted to examine the value of artificial neural networks to support the diagnosis of metabolic syndrome in a schizophrenic population. Since patients with schizophrenia are at risk to develop metabolic syndrome, a non-invasive screening technique based on a few simple clinical parameters would be very useful. Linear classification techniques are insufficient to solve this problem, but other research groups have reported favourable results using artificial neural networks. Methods: Data was acquired in the Universitair Psychiatrisch Centrum Kortenberg, on a group of 605 schizophrenic patients. First attributes were selected using several approaches using the J48 algorithm in Weka. Then different neural network models were examined in R. This led to the selection of multi-layer perceptrons with varying numbers of hidden neurons. In a third phase, 3840 neural network models were fitted on the data using repeated instances of x-fold crossvalidation. Performance on unseen data for all models was used to assign weights for the next model building step. In the fourth phase, three combination models were constructed, aimed for high sensitivity, high specificity and high PCC, respectively. Combinations were tested on the Kortenberg set, which was mixed with a random Gaussian error. Results: Increasing the number of hidden neurons from 1 to 3 improved both the true positive and true negative rates. Increasing the number of hidden neurons from 3 to 10 further improved the true negative but not the true positive rate. Percentage correctly classified was 94% in the best performing neural network (10 hidden neurons), and 63% in the worst performing network(1 hidden neuron). Performance of the three combination models was similar when they were submitted to the same noise levels. Gaussian noise with a mean of 0 and a standard deviation of 0.2 resulted in a PCC of 78%, sensitivity of 74%, and sensitivity of 80% for all three combination models, while Gaussian noise with a mean of 0 and a standard deviation of 1 resulted in a PCC of 70%, sensitivity of 60% to 61% and specificity of 75%. Conclusion: In conclusion, this study confirms the value of neural network classifiers to support the diagnosis of metabolic syndrome in schizophrenia, based on simple clinical parameters. Our findings also illustrate the importance of following a statistically correct evaluation procedure for selecting attributes, scaling the data, and then selecting, fitting, and evaluating the model.