Discretionary lane-changing (DLC) is a complex and safety-critical maneuver in highway driving that presents substantial challenges to traffic safety and operational stability. Accurately modeling DLC decisions remains difficult. Many existing approaches have limitations. They often fail to capture dynamic risk interactions with surrounding vehicles. They also inadequately model the temporal evolution of driver intent. These limitations often lead to limited prediction accuracy and generalizability. To address these gaps, we propose a new model, the Spatiotemporal Risk Field-integrated DLC (SRF-DLC) model. This model integrates time-sequential vehicle kinematics, driving behavior, and risk representation using an Attention-based LSTM network. The model was trained on trajectories from the highD dataset. We used 5,631 trajectories, each lasting 1 to 4 seconds. The model forecasts three maneuvers: lane-keeping, overtaking, and fold-down. Results show that the model trained on 3-second trajectories performs best. It achieves a peak accuracy of 96.74%. This optimal history length aligns with natural human decision-making timelines. Ablation studies were also conducted. By that we confirm the critical role of the Spatiotemporal Risk Field. Excluding this component caused a significant drop in accuracy. The highest accuracy decrease reached 47%. In comparisons, the SRF-DLC model outperformed some existing methods. This demonstrates that explicit spatiotemporal risk contextualization and appropriate history length selection is crucial in accurate DLC prediction.