The Label Is Not the Disease: Recognition Delay in Diagnosis Codes and Its Cost to EHR Case Finding
Proceedings of the Machine Learning for Health Symposium (ML4H 2026), PMLR 346, 2026
Accepted to the ML4H 2026 Proceedings track; to appear December 2026
Among UK Biobank participants whose enrollment blood assays were in the disease range for diabetes, chronic kidney disease, or anemia, a large share had no diagnosis or treatment record, recognition in primary-care and hospital records took years, and case-finding models trained on recorded diagnoses found fewer unrecognized cases than models trained on the measured assays.