JMIR Publications

ABSTRACT

Background:

Electronic health record (EHR) research increasingly identifies potentially underdocumented chronic disease by applying laboratory-based case definitions and determining whether a corresponding diagnosis code appears in the available EHR. These operational definitions measure differences between lab evidence and recorded diagnoses, but do not by themselves establish that a clinical diagnosis was warranted or missed. Their sensitivity to case-definition choices and EHR data-capture structure, and the consequences for predictive models trained on these labels, remain incompletely characterized within a single cohort.

Objective:

To characterize how EHR underdocumentation for hypertension (HTN), chronic kidney disease (CKD), and type 2 diabetes mellitus (T2DM) varies with case-definition strictness, health care engagement, and sociodemographics; to also test whether classifiers trained on lab-positive/no-code labels depend on the case definition and on observation-process features unrelated to disease severity.

Methods:

Cross-sectional analysis of All of Us Research Program participants. Cases were defined using laboratory-based operational criteria for T2DM and CKD and an HTN phenotype requiring at least 2 elevated blood pressure days at ≥130/80 mm Hg, using the threshold introduced in the 2017 ACC/AHA guideline. We trained gradient-boosted classifiers in two experiments. First, we compared models trained using weak versus sustained HTN labels. Second, we compared models using disease-severity, observation-process, demographic, or combined feature sets. The observation-process and demographic feature analyses were repeated for T2DM and CKD.

Results:

Among 226,593 lab-positive participants, baseline no-code fractions were 8.0% for T2DM, 45.5% for HTN, and 19.0% for CKD. Stricter HTN definitions reduced the fraction to 29.6%, which remained substantial. Underdocumentation was more prevalent in milder disease and persisted among heavily engaged patients. For HTN and CKD, no-code fractions were higher among privately insured and higher-income participants. Weak- and sustained-label HTN models produced near-identical ranking. For HTN, both disease-severity and demographic features predicted no-code status better than observation-process features, which added independent signal beyond disease severity. For T2DM and CKD, observation-process features substantially outperformed demographic features.

Conclusions:

Operational definitions and data-capture structure materially shape EHR underdocumentation. At the 130/80 mm Hg hypertension threshold, the apparent HTN underdocumentation rate was substantial and remained so under increasingly strict persistence requirements. Prevalence estimates were definition-sensitive, whereas patient rankings were highly stable across the two HTN definitions compared in the modeling experiment. Classifiers trained on these labels capture substantial signal from features unrelated to disease severity, including in conditions with relatively complete coding. Research and clinical ML using such labels should report sensitivity to operational choice and interpret predictions as documentation status within the available EHR extract, not missed clinical diagnosis.