A potential drawback of using unidimensional seismic demand models conditioned on ground motion intensity measure is their limited ability to efficiently evaluate the effects of changes in deterioration parameters or structural and geometric parameters variation on bridge seismic demand. Consequently, this study develops multidimensional seismic demand models for aging highway bridge class using modern machine learning algorithms. The seismic demand models for the highway bridge class in this study are conditioned on structural modeling parameters, deterioration parameters, and geometric parameters along with ground motion intensity measure. For the aging highway bridge class, a detailed finite element (FE) model is developed, and nonlinear time-history simulations are carried out considering the range of modeling, deterioration, and geometric parameters. Different machine learning (ML) algorithms are used to fit the obtained bridge component responses as a function of earthquake intensity and predictor variables. Goodness-of-fit measures are utilized to evaluate the predictive capability of the adopted models. The findings from this study reveal that extreme gradient boosting (XGBoost) and random forest (RF) performance are significantly better than the polynomial response surface model (PRSM). Moreover, the extreme gradient boosting algorithm results in the most accurate predictions with a high adjusted R2 value, low root mean square error, and a lower symmetric mean absolute percentage error for various bridge component responses.