To address the challenges of high uncertainty in sample data and limited prediction accuracy in the field of seismic damage prediction of buildings, this study proposes a multi-model fusion method based on Dempster-Shafer (D-S) evidence theory for stratified prediction. First, a seismic damage database was established, and six machine learning models—Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Self-Attention and Intersample Attention Transformer (SAINT), Tabular Network (TabNet), and Tabular Transformer (TabTransformer)—were optimized for training. Submodels were selected through performance comparisons and heterogeneous combination strategies. Second, based on the characteristics of the submodels, a modified probability distribution function was constructed. D-S evidence theory was applied to calculate conflict coefficients and confidence interval widths, converting the uncertainty of sample data into quantifiable decision risk values. Differentiated fusion strategies were implemented based on risk levels. Finally, the proposed method was validated using the 2015 Nepal earthquake database, which includes 762,106 building samples. The results show that in the unstratified prediction scenario, the overall prediction accuracy of the sample is 53.70%, an improvement of 2.21% over the optimal single model XGBoost. In the stratified prediction scenario, the prediction accuracy of the deterministic and uncertain samples increases to 67.91% and 56.41%, respectively, representing relative increases of 29.25% and 7.35% compared to the optimal single model accuracy. The example verifies the effectiveness of this method in quantifying data uncertainty and improving prediction accuracy and decision reliability.