Mining Electronic Health Records Using Artificial Intelligence: Bibliometric and Content Analyses for Current Research Status and Product Conversion
Comprehensive Overview
Published in the Journal of Biomedical Informatics in 2023, this study asked how the growing field of using artificial intelligence to mine medical information from electronic health records has actually developed over the past decade, and, critically, how much of that research has translated into real tools doctors and patients can use. Researchers systematically searched five major databases, including PubMed and Web of Science, for studies published between January 2009 and July 2022 that used AI methods such as machine learning to extract medical information from EHR data. Following PRISMA screening guidelines, they narrowed an initial 2,326 records down to 631 qualifying articles involving 4,020 authors across 267 journals, then combined bibliometric analysis, examining publication trends, leading countries, journals, and research topics, with content analysis of the AI techniques, medical data types, and clinical findings each study reported, and a novel assessment of how often the resulting medical knowledge was converted into actual products like clinical guidelines or software. The topline result was that this research area has grown rapidly since 2017, dominated by classification-based deep learning approaches, but only about one in five funded studies produced a tangible applied product.
Outcomes and Implications
The single most important finding was that, despite explosive growth in this research field, averaging 55.7% annual publication growth after 2017 and peaking at 171 articles in 2021, translation of findings into practical use remains strikingly limited: only 127 of the 631 studies, about 20%, produced an applied product such as a clinical guideline, medical system, or piece of software, and among the 437 funded studies, just 92 yielded a usable product, for an overall input-output conversion rate of only 21.05%. Supporting this picture, the field is heavily concentrated geographically, with the United States and China together producing most publications yet showing little direct international collaboration between them. Methodologically, classification tasks dominated (66% of studies), most often tackled using deep learning, decision tree, and regression algorithm families, while research topics have gradually shifted from simple disease risk grading toward more sophisticated disease risk prediction and automated extraction of disease features from clinical text and images. Data access remains a bottleneck, as only 20% of studies used openly available EHR datasets, with most researchers relying on proprietary data from their own or partnering institutions. On a more encouraging note, medical professionals were directly involved as lead or corresponding authors in 68% of studies, suggesting clinical relevance is generally prioritized. The authors conclude that while this field is maturing quickly, future progress depends on strengthening international collaboration, improving open access to EHR data, prioritizing interpretable AI algorithms clinicians can trust, and incentivizing the conversion of findings into deployable clinical products.