This study presents a novel Machine Learning (ML) approach for predicting the central and minimum film thickness in isothermal elastohydrodynamically lubricated (EHL) smooth line contacts under non-Newtonian conditions. A modular, generalized ML approach based on the similarity analysis is proposed and compared with a direct one. The Moes dimensionless parameters and pressure-viscosity adaptation are employed to account for different operating conditions, solid body, and lubricant properties, incorporating non-Newtonian effects and correction for compressibility. Both ML approaches are trained using a database generated by solving the coupled generalized Reynolds, linear elasticity, load balance, and shear stress equations for EHL line contacts using a Finite Element Method (FEM)-based numerical model. The database comprises 1055 data points for Newtonian conditions and 1004 data points for non-Newtonian conditions. The widely used ML regression algorithms are implemented comprehensively for both approaches. Prediction accuracy and robustness are then analyzed in detail for the best-performing ML model configurations. The modular approach achieves a slightly lower predictive accuracy, but enhanced robustness, compared to the direct approach. Moreover, it offers greater flexibility in terms of decoupling physical effects and reducing dataset requirements and provides good generalization capability. This could potentially extend the proposed modular approach to thermal EHL, and incorporate other non-conventional configurations (e.g., surface coatings and mixed lubrication).