Prediction of clinical scores from neuroimaging data with censored likelihood gaussian processes

In this paper, we explore the use of Censored Likelihoods in Gaussian Process Regression when predicting bounded clinical scores from neuroimaging data. The standard approach, which uses a Gaussian Likelihood, does not respect the fact that the clinical scores are bounded, and so may produce suboptimal models. Conversely, Censored Likelihoods explicitly model the restricted range of such clinical scores and carry this property through inference. We apply both the standard approach and the Censored Likelihood approach to the prediction of the MMSE score from structural MRI. Overall, we find small improvements in mean squared error when using the Censored Likelihood and in addition, the censored models are more favoured from a Bayesian perspective. We also discuss the qualitative nature of the predictions of the two approaches.