Railway track defects impact ride quality, safety, and system functionality, serving as key indicators of structural health. While superstructure defects are easily measurable, substructure components are concealed and less homogeneous, making assessment challenging. Substructure stiffness critically affects track dynamics and support conditions, and accurate measurement enhances predictive maintenance decisions. This study employs a data-driven approach to estimate railway track stiffness and curvature using vehicle response data. A railway vehicle-track system, including superstructure and substructure elements, was modeled in Simpack®. The model incorporated realistic track stiffness variations due to ballast and subgrade damage. Simulated data included sensor measurements from the axle box, bogie, and car body aligned with actual instrumented railway vehicles (IRVs) and those found in the literature. Six regression machine learning (ML) models—ExtraTrees, XGBoost, Gradient Boosting, CatBoost, Random Forest, and AdaBoost—were used for track stiffness estimation. Feature engineering, selection, and hyperparameter optimization were performed to enhance model accuracy and optimize sensor configuration. The effective sensor placement for predicting track stiffness variation was identified as accelerometers positioned on the front car body and axle box of the rear bogie’s leading wheelset, utilizing 10 statistical metric features for precise predictions. The ExtraTreesRegressor emerged as the best ML model, achieving an R2 of 0.9882, with low root mean square error (RMSE) magnitude relative to data dispersion. Applying a 4th-order Butterworth low-pass filter improved the RMSE by 33.2%, enhancing prediction accuracy. Furthermore, the ExtraTreesRegressor model effectively predicted track curvature variation using the same sensor configuration, achieving an R2 of 0.9891 without requiring additional feature engineering.