Abstract: Artificial intelligence and data-driven EHR mining can discover potential medical laws and knowledge, and provide high-value intelligence and technical method support for precise and personalised medical decision-making and health management. In this paper, we retrieve relevant EHR mining literature from Web of Science, PubMed, and CNKI databases, and analyse the hotspots and trends of research in the field by visualising the trend of publication and keyword co-occurrence. On the basis of a full understanding of EHR data types and database sources, the existing EHR mining technology methods in the scientific community and their advantages and disadvantages are summarised and compared and analysed. The current EHR mining techniques can be divided into four types: association rule-based, dictionary and rule combination, statistical machine learning, and deep learning. Among them, deep learning-based EHR data mining technology is the current research hotspot and trend, which can efficiently mine and predict large-scale complex and heterogeneous EHR data. The overall research still exists problems such as poor interpretability of mining results, single and insufficient integration of technical methods, low degree of intelligence and poor portability, poor representation learning ability of multimodal heterogeneous data, and difficulties in landing practical applications in the medical field. Future research should focus on the interpretability of EHR mining results, strong representation of multimodal heterogeneous data, integration and standardisation of EHR data, and landability in clinical medical practice. In addition, with the rapid development of technologies related to large language models and knowledge graphs, explore the feasibility of their practical application in the field of EHR mining.

Key words: Electronic Health Records (EHR), Electronic Medical Records, Data Mining, Machine Learning, Deep Learning