ABSTRACT OBJECTIVE To evaluate the accuracy of yellow fever (YF) suspected case definitions from the Brazilian Ministry of Health (BMH) and World Health Organization (WHO), as well as propose and evaluate new definitions of suspected cases, considering confirmed and discarded cases. METHODS The retrospective study was conducted at the Instituto de Infectologia Emílio Ribas (IIER), using the Epidemiologic Surveillance Form of YF cases. From the confirmed and discarded cases of YF, a logistic regression model was developed. The independent variables were used in a proposed definition of a suspected case of YF and its accuracy was evaluated. RESULTS In total, 113 YF suspect cases were reported, with 78 confirmed (69.0%). The definitions by BMH and WHO presented low sensitivity, 59% and 53.8%, and reduced accuracy, 53.1% and 47.8%, respectively. Predictive factors for YF were thrombocytopenia, leukopenia, and elevation of transaminases greater than twice normal. The definition including individual with acute onset of fever, followed by elevation of ALT or AST greater than twice the reference value AND leukopenia OR thrombocytopenia presented high sensitivity (88.3%), specificity (62.9%), and the best accuracy (80.4%), as proposed in the model. CONCLUSION The YF suspected case definitions of the BMH and the WHO have low sensitivity. The inclusion of nonspecific laboratory tests increases the accuracy of YF definition.