Existing research efforts into the decision-making of intelligent vehicles mainly focused on crash avoidance, yet have not considered the potential of injury mitigation under safety-critical scenarios. Intending to minimize occupant injury risks, this study presents a prediction-uncertainty-aware safety decision-making algorithm that takes real-time predicted injury information with quantified prediction uncertainties as decision reference. The proposed algorithm makes decisions via a periodically iterative optimization method, which can instruct vehicles to find the optimal collision configuration with minimal occupant injuries when confronting an impending collision. To enhance its robustness, two kinds of uncertainty are considered when making such safety-critical decisions: occupant injury uncertainty for the ego vehicle and driving intention uncertainty for surrounding vehicles. Simulation-based experiments with ablations are carried out to validate its ability to avoid crashes or, if the collision was inevitable, mitigate occupant injury risks. The proposed algorithm is anticipated to improve existing active safety systems under safety-critical scenarios and eventually enhance road traffic safety.