Abstract
<jats:p>Machine learning research in intelligent transportation systems is still largely evaluated through predictive accuracy, often treated as the primary indicator of model quality and practical value. Although this emphasis has supported substantial technical progress, it overlooks critical human and societal concerns, including interpretability, unequal impacts across user groups, uncertainty in high-stakes decisions, and the limited ability of practitioners to translate predictions into safe interventions. This Perspective proposes a human-centric decision-support framework for machine learning in safer intelligent transportation systems. The framework retains task-appropriate predictive performance as a prerequisite and complements it with six human-centric dimensions: relevance, explainability, fairness, uncertainty, human oversight, and actionability. Together, these dimensions provide a basis for evaluating whether predictive systems are not only accurate, but also understandable, trustworthy, and usable in real operational contexts. The framework has direct implications for researchers designing transport models, municipalities deploying data-driven safety policies, and transport operators integrating automated recommendations into daily decision-making. It argues that progress in intelligent mobility should be measured not only by prediction quality, but also by the quality of the decisions that predictive systems enable.</jats:p>