Purpose/Significance The relative positions of causality words are utilized to assist deep learning models to improve causality prediction and mine medical text gain information.Method/Process The relative position information of causality words in medical texts is represented as a relational feature layer embedded in a pre-trained language model, and the baseline model is integrated for entity recognition and relationship extraction.Result/Conclusion The F1 value of the model embedded in the relational feature layer is improved by 2.92 percentage points and 6.41 percentage points compared with the baseline models BERT-BiLSTM-CRF and CasRel, respectively, with better causal prediction capacity.