A Bayesian Algorithm for Image-Based Time-to-Event Prediction

This paper presents a novel Bayesian algorithm for making image-based predictions of the timing of a clinical event, such as the diagnosis of disease or death. We build on the Relevance Voxel Machine (RVoxM) framework, a Bayesian multivariate prediction model that exploits the spatial smoothness in images and has been demonstrated to offer excellent predictive accuracy for clinical variables. We utilize the classical survival analysis approach to model the dynamic risk of the event of interest, while accounting for the limited follow-up-time, i.e. censoring of the training data. We instantiate the proposed algorithm (RVoxM-S) to analyze cortical thickness maps derived from structural brain Magnetic Resonance Imaging (MRI) data. We train RVoxM-S to make predictions about the timing of the conversion from Mild Cognitive Impairment (MCI) status to clinical dementia of the Alzheimer type (or AD). Our experiments demonstrate that RVoxM-S is significantly better at identifying subjects at high risk of conversion to AD over the next two years, compared to a binary classification algorithm trained to discriminate converters versus non-converters.