Multistate Models: Accurate and Dynamic Methods to Improve Predictions of Thrombotic Risk in Patients with Cancer.

Abstract

Research into cancer-associated thrombosis (CAT) entails managing dynamic data that pose an analytical challenge. Thus, methods that assume proportional hazards to investigate prognosis entail a risk of misinterpreting or overlooking key traits or time-varying effects. We examined the AGAMENON registry, which collects data from 2,129 patients with advanced gastric cancer. An accelerated failure time (AFT) multistate model and flexible competing risks regression were used to scrutinize the time-varying effect of CAT, as well as to estimate how covariates dynamically predict cumulative incidence. The AFT model revealed that thrombosis shortened progression-free survival and overall survival with adjusted time ratios of 0.72 and 0.56, respectively. Nevertheless, its prognostic effect was nonproportional and disappeared over time if the subject managed to survive long enough. CAT that occurred later had a more pronounced prognostic effect. In the flexible competing risks model, multiple covariates were seen to have significant time-varying effects on the cumulative incidence of CAT (Khorana score, secondary thromboprophylaxis, high tumor burden, and cisplatin-containing regimen), whereas other predictors exerted a constant effect (signet ring cells and primary thromboprophylaxis). The model that assumes proportional hazards was incapable of capturing the effect of these covariates and predicted the cumulative incidence in a biased way. This study evinces that flexible and multistate models are a useful and innovative method to describe the dynamic effect of variables associated with CAT and should be more widely used.