A Systematic Approach to Prioritise Diagnostically Useful Findings for Inclusion in Electronic Health Records as Discrete Data to Improve Clinical Artificial Intelligence Tools and Genomic Research.

Aims: The recent widespread use of electronic health records (EHRs) has opened the possibility for innumerable artificial intelligence (AI) tools to aid in genomics, phenomics, and other research, as well as disease prevention, diagnosis, and therapy. Unfortunately, much of the data contained in EHRs are not optimally structured for even the most sophisticated AI approaches. There are very few published efforts investigating methods for recording discrete data in EHRs that would not slow current clinical workflows or ways to prioritise patient characteristics worth recording. Here, we propose an approach to identify and prioritise findings (phenotypes) useful for differentiating diseases, with an initial focus on relatively common small B-cell lymphomas.

Materials and methods: A website enabling crowd-sourced recording of diseases and phenotypes was developed. An expert committee in the field of B-cell lymphomas standardised phenotype terminology for use in digital resources, and select terms were included in the Human Phenotype Ontology (HPO). A total of 100 patient lymph node biopsy samples were evaluated, and phenotypes were recorded as discrete data. Bayesian networks (BNs) were developed based on these data, and their diagnostic accuracy and ability to prioritise these phenotypes for inclusion in EHRs were assessed.

Results: Out of 146 phenotypes identified from the website as potentially useful for differentiating four different lymphomas from each other and from benign lymph nodes, 70-75 were included in BNs. The diagnostic accuracy of different naïve BNs was 96.3% for non-marginal zone lymphoma cases and 50% for marginal zone lymphoma cases when all of the included phenotypes were used and 93.8% for non-marginal zone lymphoma cases and 27.5% for marginal zone lymphoma cases when only 15 phenotypes were included in the BNs.

Conclusion: This pilot provides a starting point for systematic improvement and a dataset for comparing related approaches.