Traffic accidents pose a significant societal issue. Different strategies can be implemented to improve road safety and reduce the negative impact of traffic accidents. One such approach is driving intention recognition, which aims to predict driving intentions and alert drivers in risky situations.This study examined how temporal distance to a roundabout affects the performance of LSTM (Long Short-Term Memory) machine learning models in predicting driving intentions, specifically, whether light vehicles will turn right or continue straight at a roundabout. The dataset was collected from a roundabout in Gothenburg. Five LSTM models were developed using data points with temporal thresholds at 0, 1, 3, 5, and 10 seconds before vehicles entered the roundabout. All models were trained on the five different subsets from the same dataset by implementing a Long Short-Term Memory architecture and evaluated using the F1-score evaluation metric. The results indicate that temporal distance significantly influences model performance, with F1 scores decreasing as the temporal distance to the roundabout increases. The model trained to predict turns at the point of exit (0 seconds before the roundabout) achieved an F1 score of 0.87. In contrast, the model predicting turns 10 seconds before the roundabout exit produced a reduced F1 score of 0.75. These findings provided a better understanding of intention recognition systems and their potential role in improving traffic safety and reducing accident risk. Future research should explore how spatial distance, in addition to temporal distance, affects model performance and predictive accuracy.