Active Side Guard Assist (ASGA) plays an important role in enhancing urban traffic safety, particularly by mitigating potential collisions with Vulnerable Road User (VRU) during turning maneuvers. While current ASGA implementation has demonstrated strong safety performance, the algorithm is based on manually defined, rule-based logic. Rule-based systems require extensive effort to adapt to complex, dynamic real-world environments and may lead to false positive braking interventions. To explore potential enhancements, this thesis proposes a Machine Learning (ML)-based framework for braking decision optimization in ASGA. A structured literature review was conducted to analyze the existing methods for tuning vehicle parameters and to identify opportunities for greater adaptability and precision. Building on these insights, several supervised ML classifiers were developed and evaluated to distinguish between critical and non-critical braking events. Among these, XGBoost achieved the highest precision (97%), demonstrating strong generalization to previously unseen scenarios. Comparative analysis further underscored the advantages of ML-based approaches over traditional rule-based methods in terms of precision and adaptability. The integration of Explainable AI (XAI) techniques and the expert feedback further supported transparency and interpretability. This thesis demonstrates that data-driven methods can further improve ASGA performance by supporting the development of safer, more adaptive active safety system and aligning with the long-term vision of achieving zero traffic fatalities.