From theory to practice - assessing translation of physical fitness research in the emergency department through machine learning and natural language processing.

Background: A critical challenge for biomedical investigators is the delay between research and its adoption, yet there are few tools that use bibliometrics and artificial intelligence to address this translational gap. We built a tool to quantify translation of clinical investigation using novel approaches to identify themes in published clinical trials from PubMed and their appearance in the natural language elements of the electronic health record (EHR).

Methods: As a use case, we selected the translation of known health effects of exercise for heart disease, as found in published clinical trials, with the appearance of these themes in the EHR of heart disease patients seen in an emergency department (ED). We present a self-supervised framework that quantifies semantic similarity of themes within the EHR.

Results: We found that 12.7% of the clinical trial abstracts dataset recommended aerobic exercise or strength training. Of the ED treatment plans, 19.2% related to heart disease. Of these, the treatment plans that included heart disease identified aerobic exercise or strength training only 0.34% of the time. Treatment plans from the overall ED dataset mentioned aerobic exercise or strength training less than 5% of the time.

Conclusions: Having access to publicly available clinical research and associated EHR data, including clinician notes and after-visit summaries, provided a unique opportunity to assess the adoption of clinical research in medical practice. This approach can be used for a variety of clinical conditions, and if assessed over time could measure implementation effectiveness of quality improvement strategies and clinical guidelines.