This research proposes a statistical confidence interval-based model refinement approach for the health diagnosis of structural systems subjected to seismic excitations. The proposed approach uses the 95% confidence interval of the estimated structural parameters to determine their statistical significance in a least-squares regression setting. When the confidence interval of the parameters covers the ”null” value, it is statistically sustainable to truncate them. Thus, the remaining parameters repetitively undergo such a sifting process for model refinement until all statistical significance cannot be further improved. Other confidence intervals, such as the 90% and 99%, of structural parameters are also tested for comparison and validation purposes. This newly developed model-refinement approach is implemented for the developed series models of multivariable polynomial expansions: the linear, as well as the Taylor and the power series models, thereby leading to a more accurate identification of structures for safety assessment.