The study drew on 2,270 health-examination records generated in 2025 through China’s National Basic Public Health Service program, which provides annual checkups for residents aged 65 and older. The records came from a rural primary-care program in Wangkui County, Heilongjiang Province, an agricultural region where 20 examination sites contributed data. The researchers, led by Kun Xu of Henan University’s Institute of Gerontology, set out to answer a deceptively simple question: could the routine main-medication field in these records be trusted as a measure of medication exposure in secondary research?

The answer turned on a comparison that initially looked compelling. Records were split by the exercise-frequency item into those with recorded physical activity, meaning daily, weekly, or occasional exercise, and those recording no exercise at all. Among the 2,226 records with usable activity information, nonblank medication fields appeared in 56.4 percent of the active group versus 42.8 percent of the inactive group, a crude difference of 13.6 percentage points. A modified Poisson regression, chosen because the outcome was common, produced a prevalence ratio of 1.317 with a confidence interval comfortably excluding the null. On its face, older adults who exercised appeared far more likely to be taking medications.

But that interpretation collapses under scrutiny of the data-generating process. Under the national health-service specification, the main-medication item was never designed as a population-wide drug inventory. It was intended to document principal medications used during the preceding year among people receiving long-term treatment for chronic disease, specifically within hypertension and diabetes management programs. When the researchers cross-tabulated field presence against chronic-disease management registration, the pattern was stark: 96.3 percent of records with hypertension and/or diabetes registration had nonblank medication fields, compared with just 2.4 percent of records with neither registration. The field was essentially a switch flipped by enrollment in a disease-management pathway.

Because exercise status was itself correlated with registration, the apparent activity contrast was a mixture of two things: any real relationship between activity and medication use, and the fact that active records were more likely to belong to the registered group where medication fields were almost always filled. The researchers built a sequence of models to watch the contrast dissolve. After adjusting for age, sex, and examination site, the prevalence ratio actually rose slightly to 1.346. But adding a four-category hypertension-diabetes registration variable, which allowed each combination of registrations its own effect, dropped the ratio to 1.001, with a confidence interval from 0.960 to 1.043. Further adjustment for body mass index, smoking, and alcohol use left it essentially unchanged at 1.002.

The absolute scale told the same story. Marginal standardization of a logistic model, which sets every record’s activity status in turn while holding all other covariates at observed values, yielded standardized field-presence proportions of 51.26 percent for the active group and 51.14 percent for the inactive group, a difference of just 0.12 percentage points with a confidence interval spanning negative 2.07 to 2.30. Within registration strata, the differences were tiny: negative 1.41 percentage points among unregistered records, 0.44 among hypertension-only records, and negative 0.56 among dual-registered records. Only the sparse diabetes-only stratum, with just 37 records, showed a large descriptive difference, and the authors explicitly declined to generalize from it.

The team did not stop at field presence. They audited the medication text itself, splitting 1,141 nonblank fields into 1,368 retained text entries after removing exact duplicates. A version-controlled rule-based parser identified medication names, doses, units, and duration information, and a random sample of 200 entries was independently reviewed by two pharmacists masked to the parser’s output. Inter-reviewer agreement on broad therapeutic class reached 99.5 percent, with a kappa of 0.982, and parser-to-manual concordance on that domain was 99.0 percent. Archived labels classified 75.9 percent of entries as antihypertensive and 14.5 percent as glucose-lowering, with 7.4 percent unresolved. Yet the authors are careful to note that this high concordance validates only the interpretation of recorded text, not actual medication exposure, because no linked prescriptions, dispensing records, or medication reconciliation data existed.

The audit also exposed structural ambiguities that would plague any secondary analysis. Nearly half of the entries, 45.0 percent, contained a bare number in the duration field with no time unit, and another 15.5 percent held timing or frequency information rather than a treatment duration. Standardized medication naming showed more variability than direct text extraction, with parser-to-reference concordance of 77.5 percent on exact normalized matching, reflecting differences in formulation qualifiers, salt wording, and dictionary granularity. A blank field, meanwhile, could mean no long-term medication, inapplicability, the question never asked, unknown status, refusal, or simple omission, all compressed into an undifferentiated empty cell.

The implications reach well beyond one county in Heilongjiang. Routine health data are products of healthcare and administrative processes rather than neutral research measurements, and informative field presence can generate associations that reflect documentation architecture rather than clinical reality. The authors sketch the specific risks: coding blank fields as medication nonuse would produce differential exposure misclassification patterned by disease registration; restricting analyses to records with nonblank fields would introduce selection bias by preferentially retaining records from chronic-disease pathways; and the intended one-year reference period may misalign with outcomes measured at a single visit. They propose a tiered approach, reserving the field for documentation audits and narrowly defined descriptive work within its intended population, while requiring validation against prescriptions, dispensing data, or structured medication reconciliation for any study of drug utilization, safety, effectiveness, or medication-outcome associations.

The study also offers concrete fixes for primary-care information systems: separating disease-management prompts from comprehensive medication lists, storing identity, indication, dose, route, frequency, and status in distinct structured elements, adopting standardized terminologies such as RxNorm and the ATC/DDD classification, and replacing undifferentiated blanks with explicit categories distinguishing nonuse, inapplicability, and missingness. The limitations are acknowledged candidly, including the single-center setting, the cross-sectional design, the absence of stable person-level identifiers, and reliance on only 20 site clusters. But the central message stands with unusual clarity: before any routinely collected variable is reused as research data, its intended population, recording workflow, and reference period must be understood. A field fit for its original operational purpose can be quietly unfit for the next question asked of it, and the ghost of that purpose can masquerade as a scientific finding.

Subject of Research: Secondary-use suitability of a routine primary-care medication field and its association with recorded physical activity among rural older adults

Article Title: Among rural older adults, an apparent physical-activity contrast in nonblank medication-field presence was markedly attenuated when chronic-disease management registration was incorporated into the analysis

Article References: Among rural older adults, an apparent physical-activity contrast in nonblank medication-field presence was markedly attenuated when chronic-disease management registration was incorporated into the analysis. (n.d.). https://doi.org/10.3389/fphar.2026.1965382

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1965382

Keywords: pharmacoepidemiology, real-world data, medication data quality, electronic health records, older adults, physical activity, chronic disease management, exposure misclassification, primary care, rural China, data provenance, selection bias