Prediction and accuracy of risk conversion from MCI to probable Alzheimer's disease using longitudinal-survival joint modeling in ADNI datasets

Abstract

In many cohorts or clinical studies looking at biomarker evolution, there are two types of data of interest: longitudinal (e.g. hippocampal volumes (HC) from MRI, FDG-PET metabolism, CSF ABeta 1-42) and time-to-event (e.g. time of conversion from mild cognitive impairment (MCI) to probable Alzheimer's disease (AD)). It is hypothesized that the evolution of these biomarkers is associated with disease severity and thus progression from MCI to AD. In this work, we wished to study this progression, taking into account both the evolution and the risk progression, by combining a longitudinal model for HC as a function of Mini-Mental State Evaluation (MMSE) and a survival model of conversion from MCI to AD. The resulting model is a novel, joint model for longitudinal and survival data. We selected a total of 121 subjects (43 F, 78 M) having progressed from MCI to AD in the ADNI, who had at least two visits within a 36 months follow-up. Due the nature of joint modeling methodologies, we used MMSE as a surrogate of time however, because the time variable must be strictly monotonic, we replace MMSE by its predicted value fitted from a linear model. A mixed model was used to fit HC and a Cox model to fit the conversion risk. The joint model was obtained for every individual by combining these two models. To estimate the global association between HC evolution and conversion risk, the joint model was fitted and applied to a new fictitious “average”-subject, represented as the average of all the subjects. Thus, the probability of conversion and its accuracy for the whole group were estimated using this technique. All statistical analyses were done using the R and JM software packages. Demographic information can be found in table 1. The probabilities for global risk conversion are shown in Figure 1, and predictive accuracy in Figure 2. Our results show on average that as HC decreases the risk of conversion increases, as expected. Prediction of global probability of risk conversion from MCI to AD for converted subjects. As HC decreases, the risk of conversion increases. ROC curves for measuring the predictive accracy of risk conversion from MCI to AD. This methodology can be improved using any other biomarkers; it can further serve to determine individual risk prediction of conversion in a clinical setting.