The seismic performance of vertically irregular buildings remains a critical challenge due to the complex interactions among stiffness, mass, and strength irregularities, which often lead to instability-driven failures. Conventional drift- and force-based assessment methods are insufficient to capture collapse-related mechanisms, particularly under significant second-order (P–Δ) effects. This study proposes a novel stability-driven predictive framework that integrates nonlinear structural analysis with machine learning to explicitly evaluate structural stability. A dataset comprising 58 structural configurations under three material scenarios (totaling 174 simulation cases) was generated using nonlinear time-history analysis. Key stability indicators—including the Global Stability Index (GSI), normalized P–Δ ratios, and stiffness degradation measures—were extracted and used as predictive features. Among the tested models, XGBoost achieved the highest performance (R2 = 0.91, RMSE = 0.26, accuracy = 87%), demonstrating superior capability in capturing nonlinear stability behavior. The results confirm that stability-based indicators provide significantly improved predictive insight compared to conventional drift measures. The main contribution of this study lies in introducing a physically interpretable, stability-oriented ML framework that enables rapid and reliable seismic performance classification of irregular buildings, supporting advanced design and decision-making processes.