Urban road traffic is particularly complex due to limited land resources, dense populations, and mixed transportation modes, leading to a higher likelihood of crashes. Traditional studies rarely combine macro influencing factors with micro dynamics to predict individual risks in real time. This paper takes into account a variety of urban road features and individual driving behavior to propose an LSTM-based real-time crash prediction method, with safety-critical events (SCEs) as surrogate measures of crashes. First, multi-source data including vehicle speed, acceleration, driving condition, weather, time of day, season, and weekday information are collected through natural driving experiments. Then, SCEs are extracted according to different urban road driving conditions, and finally, the LSTM model is constructed and evaluated. The results show that the model can predict SCE with 99% accuracy, 75% recall, 87% precision and 80% F1-score, which indicates a good predictive power. However, the relatively mediocre recall reflects the problem of dataset balance, and it is expected to be improved in subsequent studies. This study can provide an effective method for real-time risk warning and accident cause analysis.