A cross-domain model based on migration learning can solve the problems of missing annotated data and large distributional differences between domains, these are prevalent in biological data fields, like genomics, proteomics, and medical imaging. In this research paper, a novel deep transfer learning model founded on multi – source domain integration (MUCT) is put forward, building upon conventional cross – domain transfer learning approaches. Firstly, an end-to-end training mechanism is established based on deep neural networks, secondly, high-confidence target samples collected through consistency filters are trained as a way to create target domain supervisory information, and finally, the outcomes of the classification achieved among multiple source domains and the target domain are combined by means of the relative majority voting approach to enhance the model’s resilience. This approach demonstrates an excellent identification outcome for medical entities within Chinese electronic medical records, with a strict F1 value of 85.4% on the CCKS 2018 review dataset. Typical case study results validate that the migration method can effectively recognize entities such as personal information, disease symptoms, diagnosis and treatment, and drug use in patient question texts by utilizing only a small amount of annotated corpus, realizing the full utilization of existing data resources. This study provides an efficient knowledge migration paradigm for biological big data analysis, which is expected to promote the in-depth development of precision medicine and systems biology research.
