As a final barrier against the release of radioactive material, the prestressed concrete containment vessel (PCCV) plays an important role in the nuclear power plant (NPP). When subjected to seismic loads, its structural performance requires high structural integrity. During an earthquake, the PCCV experiences ground motions (GMs) consisting of horizontal and vertical components in a global coordinate system, and its structural integrity is probably overestimated if only the horizontal component is considered in numerical analyses. In this study, a novel cumulative damage index is first employed to quantitatively assess the structural integrity of PCCVs. Accordingly, the failure probability of structural integrity of the PCCV is investigated by incremental dynamic analysis (IDA) and further seismic fragility analysis considering the effect of vertical ground motions (VGMs). The results show that the effect of VGMs on structural integrity of PCCVs is gradually non-negligible with the increase of vertical intensity. Moreover, an accurate cumulative damage prediction model for the PCCV is established based on eight mainstream machine learning (ML) algorithms, and the necessity of incorporating vertical intensity measures (IMs) in ML models is investigated. The results show that the ML model considering the vertical IMs is superior in capturing the cumulative damage of the PCCVs. The best estimates are obtained by GB with R2 of 0.912, MAE of 1.301, MSE of 6.186, and RMSE value of 2.487. Finally, the SHapley Additive exPlanation (SHAP) method is adopted for interpreting the prediction results of the optimal ML model to assess the effect of each IM and its interactions on the generalization performance of the model. The proposed ML-based cumulative damage prediction model for PCCVs considering the effect of VGMs enables an accurate quantitative assessment on structural integrity of PCCVs. This provides an important reference for evaluating the potential environmental hazard due to the leaking of radioactive materials.