Lost circulation is a common problem encountered in complex formation drilling engineering, which is characterized by frequent occurrence, randomness and persistence. Accurate prediction of potential thief zone before drilling is particularly important for safe drilling. The traditional analysis of lost circulation is focused on diagnosis while drilling and summary after drilling, mainly using the means of combining engineering data and field experience, which leads to the lag of analysis results and cannot effectively guide drilling engineering design before drilling. Based on seismic attributes and lost circulation engineering data, this paper extracted the seismic attributes of drilled wells on the basis of the selection of single wells with typical lost circulation characteristics, and selected the seismic attributes with strong correlation with lost circulation prediction by time-depth relationship and adopted random forest method to identify and select the seismic attributes with strong correlation with lost circulation prediction. Then, an ensemble learning model was established by using soft voting algorithm in machine learning method. The model integrates three sub-models named logistic regression, random forest and support vector machine, and realizes the nonlinear mapping relationship between multiple seismic attributes and lost circulation engineering data and the corresponding weight characterization. At the same time, the probability of lost circulation risk distribution driven by the fusion of seismic and engineering data is obtained, and the 3D spatial distribution prediction of pre-drilling lost circulation risk layer is realized. The results show that variance, time-frequency attenuation, sweet spot and root mean square amplitude have the highest correlation with lost circulation. Combining the above attributes can achieve more accurate lost circulation risk prediction. However, excessive addition of seismic attributes cannot significantly improve the prediction accuracy, on the contrary, it will increase the calculation cost. Compared with a single machine learning model, ensemble learning model can achieve better prediction results because it combines the advantages of multiple sub-models. The practical application results show that the accuracy of lost circulation risk prediction by using seismic attributes depends on the sampling rate of seismic data. The horizontal prediction resolution of the thief zone risk is about 25 m, and the vertical prediction resolution is about 6 m (2 ms). The prediction results show that the horizontal prediction is more reliable than the vertical prediction. However, due to the influence of time-depth relationship, the longitudinal prediction accuracy may be offset. This study provides a new way to predict pre-drilling lost circulation, which is of great significance to guide well location deployment, well trajectory optimization and safe drilling.