A comparison of artificial neural network, latent class analysis and logistic regression for determining which patients benefit from a cognitive behavioural approach to treatment for non-specific low back pain

It can be difficult to select the right treatment for low back pain for a given individual. The objective of this study was to compare the use of artificial neural networks with latent class analysis and logistic regression to identify for whom a new, cognitive behavioural approach to the treatment of low back pain is indicated or contra-indicated. Data was made available to us from a cohort of low back pain patients recruited to a clinical trial of a cognitive behavioural approach. The 701 participants had at least moderately troublesome back pain of at least 6 weeks' duration and were recruited from 56 general practices in 7 regions in the UK between April 2005 and April 2007. For the purposes of this study, the main outcome measure was the Roland Morris Disability Questionnaire. We found that the artificial neural network with one hidden node and weight decay 0.1 was the optimal network for this application. The artificial neural network and the ordinary logistic regression had the lowest overall error rate, but the artificial neural network and the latent class logistic regression had superior log score. We concluded the superior log score of the techniques which can allow for nonlinear relationships between the variables suggests that these are more likely to be useful in decision support than linear models. We have shown that the artificial neural network provides the best combination of overall error rate and log score, and would be the best candidate of these three models for decision support for allocating patients to the cognitive behavioural approach to treatment of lower back pain.