A comparison of alternative approaches to dynamic predictions: An application on a short follow-up cohort of patients with Chagas disease

Prediction models have been used in several areas, especially in health science. In many situations, they can serve as a tool for clinical support, helping define risk groups by severity and assisting decision-making concerning the most appropriate treatment. A typical case is, for instance, to predict the risk of a patient dying in a specific window of time based on variables (clinical markers) recorded at the beginning of the study and over the follow-up period. This study was motivated by a clinical demand on building a death risk score for patients with Chagas disease from the SaMi-Trop prospective cohort project. Patients living in the state of Minas Gerais, Brazil, were followed for two years, and based on baseline information, a risk score was built to predict 2-year mortality. The follow-up study allowed the nature of the death risk to be dynamic. That is, clinical markers change over time, and for the two-year survivors, it is necessary to update the risk score. In this work, we compare four approaches to building dynamic scores: two naive ones that are based on updating the baseline score, another based on a new landmark, and the last one considering a joint modeling approach. Some simulation results in the literature seemed to favor joint modeling. In this paper, we compare the dynamic prediction approaches based on the Chagas disease short follow-up cohort. The predictive capacity of each risk score is assessed and compared using discrimination and calibration measures. The study results suggest that new baseline approaches (naive 1 and landmark) have better discrimination capacity and the other ones (naive 2 and joint modeling) have better calibration capacity.