9C.01: WHICH RISK FACTORS ARE IMPORTANT FOR ENDPOINT CARDIOVASCULAR EVENTS IN THE FEVER STUDY? A PRACTICAL APPLICATION OF ROC CURVES IN CLINICAL TRIALS. (June 2015)

• Record Type: Journal Article
• Authors: Zhang, X.; Zhang, Y.; Liu, L.; Zanchetti, A.
• Is Part Of: Journal of hypertension. Volume 33(2015)Supplement 1
• Publication Date: 2015-06
• Subjects: Hypertension -- Periodicals; Hypertension -- Periodicals; 616.132005
• DOI: 10.1097/01.hjh.0000467682.54801.ba
• Languages: English
• ISSNs: 1473-5598

Abstract

Abstract : Objective: Many randomized clinical trials, including the Felodipine Event Reduction (FEVER) study, have reported event reduction by BP-lowering treatment. Usually the COX model is used, and the importance of coexisting risk factors can be compared with the same model, but receiver operating characteristic (ROC) curves can also be used for analyzing relations between risk factors and endpoint events. Sensitivities and specificities of event occurrence can be calculated for randomized patients. Design and method: The FEVER study randomized 9711 Chinese hypertensive patients to more or less intense anti-hypertensive treatment during 40 month follow-up. ROC curves were drawn for FEVER data. Area under the curve (AUC) was used for the comparisons between different factors without any assumption on distributions, the ROC curves being a non-parametric method. Every risk factor of an event corresponds to a ROC curve, and its AUC is related to the risk of a given event. Software SAS9.2 was used. Results: For total cardiovascular events (TCVE), the risk factors considered were age(a), body mass index(b), screening SBP(c), screening DBP(d), smoking amount(e1) or smoking state, Y/N, (e2), serum cholesterol(f), study-end SBP(g), study-end DBP(h), left ventricular hypertrophy (i), previous diabetes (j), sex(k), previous cardiovascular disease(l). Using to different models of ROC and AUC, their risk importance sequences for TVCE were: a>h>k>l>TREAT>j>f>b>e1>i; and a>h>k>l>TREAT>j>f>e2>b>i. Using LOGISTIC regression, the sequences were: j>l>k>TREAT>i>a>d>e1>b>f; and j>l>k>e2>TREAT>i>a>d>b>f. TREAT randomization to either Felodipineor placebo Conclusions: Age is more important for TCVE than blood pressure; Both dichotomous and continuous variables may appear in the same model and the dichotomous risks are overestimated in LOGISTIC regression. ROC is more suitable than LOGISTIC regression, because ROC is a nonparametric method. A good example is the two types of smoking variables (state and amount): their risk importance is the same by ROC, but different by LOGISTIC analyses.