Based on the experimental data of concrete nonlinear creep under high stress levels (40-80% of prismatic strength), this study explores the application of machine learning methods for predicting creep deformation. A recurrent artificial neural network (ANN) and the CatBoost algorithm were employed to model the time-dependent creep strain, using stress and time as input parameters. The ANN demonstrated high predictive accuracy, with training achieving a mean square error of 0.000154, and its generated creep curves showed an excellent fit with the experimental data. In contrast, the CatBoost algorithm, while effectively capturing the physical trend that creep strain increases nonlinearly with stress and decelerates over time, exhibited lower prediction accuracy than the ANN. Feature importance analysis within the CatBoost model highlighted the significant influence of lagged stress parameters and time-squared terms, aligning with the nonlinear physical nature of concrete creep. The results confirm the strong potential of machine learning, particularly recurrent neural networks, for modeling complex nonlinear creep in concrete, even with limited datasets. Future work is suggested to incorporate concrete strength class and loading age as additional parameters to enhance model generalizability.