Excessive arousal in difficult driving scenarios is a major contributor to traffic accidents. To address this issue, this study proposes a novel method that leverages personalized neurofeedback to modulate driver arousal in real-time. Firstly, by analyzing the relationship between pupil size and reaction time, the Yerkes-Dodson law was validated. Additionally, a cascade model structure was implemented to develop an EEG-based arousal decoder, achieving an arousal recognition accuracy of 74.8%. Secondly,  the real-time generic neurofeedback was demonstrated to help a driver manage over-arousal in difficult driving scenarios. Compared to both silence and sham control conditions, the generic neurofeedback significantly extended crash-free driving time. Notably, when comparing real-time neurofeedback to silence alone, the average driving duration increased by 7.95%. Finally, instead of continuous feedback in generic neurofeedback, this study proposes a Markov decision process (MDP) framework to tailor neurofeedback strategies according to individual differences in workload and arousal. Unlike the continuous generic feedback, the MDP-based approach delivers feedback intermittently. Results indicated that the MDP-based approach further improved driving performance and stabilized arousal levels. Compared to a generic neurofeedback, the MDP strategy yielded an additional 10% increase in average driving time. The findings highlight the feasibility of incorporating EEG-derived arousal states into an adaptive neurofeedback system, confirming that personalized neurofeedback helps drivers maintain optimal arousal under difficult driving conditions, ultimately reducing driving errors and promoting safer road performance.