Modelos paramétricos de análise de sobrevivência para avaliação da evolução de pacientes renais crônicos em tratamento conservador

Abstract

Introduction: Chronic kidney disease (CKD) is associated with increased early mortality, increased cardiovascular disease (CVD), hospitalization and rehospitalizations, in addition to health complications. The patient is monitored through the estimated glomerular filtration rate (eGFR). GFR is accepted as the best global measure of kidney function. Reductions in eGFR on specific gradients are associated with a worse prognosis with consequent accelerated progression of CKD. Objective: To methodologically explore the parametric regression models in survival analysis, for proportional hazards (PH) and for accelerated failure (AFT), in the investigation of the accelerated progression of CKD (pre-dialysis phase), with surrogate outcomes, and also the incorporation random effects (fragility). Methodology: This is a methodological and critical study of two parametric regression techniques applied in a case study. We used a historical cohort with real data from predialysis patients with CKD. Interpretations and adequacy of the proposed PH and AFT models were presented for the distributions: Exponential, Weibull, Gompertz, Lognormal and Loglogistic. Models were adjusted for month until ≥ 5mL/year drop in eGFR, 30% loss in eGFR within 24 months, or change in CKD stage, as surrogate endpoints. Results: The >5mL/year decay of the eGFR presented a monotonous characteristic, in which the events decrease over time. On the other hand, the >30% decay of eGFR had an inverse relationship, with an increase in risk over time. The vast majority of stage changes occurred within 12 months (67.8%), with a tendency to be constant over time. In the inspection of the graphs, we observed that the best superposition of the curves with the empirical method of Kaplan-Meier and Smoothed Hazard Estimate occurred in the following order: A) decay >5mL/year of the GFR, indicates a better fit of the Gompertz model, for the functions S(t) and h(t). B) decay >30% of the eGFR, the best fits for the function S(t) and h(t) were the Gompertz, Weilbull and Loglogistic distributions. C) For the change of stage, the behavior of the curve is well captured by practically all models. Regarding frailty, the Weibull and Loglogistic models had better multivariate fits. With the point values of frailty at the patient level, it was possible to identify three groups sharing the same frailty values: 110 patients (14.4%) with frailty values between (1.01 - 1.99), 124 patients (16, 0%) with frailties between (2.00 – 2.99) and 46 patients (5.9%) with frailty values > 3.0. These patients have an increase in variability due to causes that have not been properly evaluated a priori. Conclusion: The more robust statistical properties of the parametric models (PM) must be taken into account, including the plausibility of the extrapolations that their results can be made. Given the flexibility of the MP, we can compare the results of the curves of the parametric S(t) and h(t) functions with relevant studies, as well as with the opinion of clinical experts in nephrology, comparing with known patterns of disease and indicating the possible ways to follow.