Generating Computable Phenotype Intersection Metadata Using the Phenoflow Library
0 likes17 views
AMIA Informatics Summit, Pittsburgh, 2025 (presented by Professor Vasa Curcin)
Generating Computable Phenotype Intersection Metadata Using the Phenoflow Library
- 2. Generating Computable Phenotype Intersection Metadata Using the Phenoflow Library Toward Implementation: Addressing Real-World Deployments S25 Martin Chapman, Vasa Curcin King’s College London Luke V. Rasmussen, Jennifer A. Pacheco Northwestern University Laura K. Wiley WashU Medicine
- 3. DISCLOSURE OF CONFLICTS OF INTEREST I have not had any relationships with ACCME-defined ineligible companies within the past 24 months.
- 4. Background: Computable phenotypes Knowledge objects that capture the logic required to identify individuals with a disease or condition from their medical records.
- 5. Phenotype libraries Online phenotype catalogues, which store a significant number of computable phenotypes for the same disease or condition.
- 6. Phenotype definition multiplicity This is a good thing (mostly)… • It is not desirable (or feasible) to have a single computable phenotype for every condition. Different use cases necessitate different logic. But… • We need to understand which use cases are already supported, to facilitate reuse. In other words, we need to understand what is unique about each phenotype. This can then be stored as metadata.
- 7. Phenotype intersection To understand what is unique about each phenotype (and thus which use cases it best supports), we can first do the opposite and understand how two phenotypes for the same condition intersect. We can aim to do this automatically and therefore at scale.
- 8. Barriers to automated intersection analysis 1. Identifying when two computable phenotypes target the same disease or condition in the first place. • e.g. ‘T2DM Implementation’ vs. ‘Type 2 Diabetes Mellitus’ (PheKB) 2. Comparing different forms of computable phenotypes • e.g. codelists vs. Natural Language Processing (NLP)
- 9. Methods: Identifying same disease/condition 1. Levenshtein distance to identify text similarity. 2. HDR UK API calls to identify phenotypes that target the same condition but lack text similarity using common keywords. 3. Large Language Model (LLM) (Llama 3.1) to validate the additional phenotypes returned in 2. (Not all 161 definitions are actually for diabetes).
- 11. Results: Intersection – Condition groups 1171 definitions loaded into the Phenoflow library. 137 condition groups (conditions with two or more phenotypes). PPV 95%. 574 definitions exist as a part of a group (49%). Good insight into the extent of the definition multiplicity phenomenon.
- 12. Results: Intersection – Steps Trend: Across the 10 largest condition groups, the average number of steps in common between pairs of definitions relative to the average number of steps in the group is low. While definition multiplicty exists, definitions still have a considerable number of unique steps.
- 13. Results: LLM impact We observed our LLM: • Identifying false positives (e.g. matches between phenotypes for different types of heart failure). • Identifying false negatives (e.g. phenotype names that do not include the condition but still aim to identify the condition via the presence of medications).
- 14. Summary and Future work The use of Phenoflow has allowed us to compare definitions to understand more about definition multiplicity (extensive) and intersection (limited). Integrating an LLM increases the reliability of this process. Unique steps will soon be added to Phenoflow as metadata to support reuse. To complement definition intersection insight (horizontal), definition subsumption (vertical) will be explored next.
- 15. Links Implementation (Python): https://github.com/phenoflow/curator Data analysis (Jupyter): https://github.com/phenoflow/intersection-analysis Live Phenoflow site: https://kclhi.org/phenoflow