Medication History Outperformed Clinical Profiles Across Three Aging Cohorts
In a multicountry analysis spanning China, Europe, and the United States, medication history proved far more informative than baseline clinical profiles for predicting who would stop taking antihypertensive therapy at the next survey wave. The finding matters because treatment non-persistence is one of the major reasons hypertension remains uncontrolled despite the wide availability of blood pressure–lowering drugs. Here, the predictive signal came not from a dense set of clinical variables, but from the patient’s own medication behavior over time: whether they were using antihypertensive therapy at baseline, how often they had been non-user in prior waves, and whether they had previously discontinued treatment.
The study harmonized three prospective aging cohorts: the China Health and Retirement Longitudinal Study, the Survey of Health, Ageing and Retirement in Europe, and the US Health and Retirement Study. Adults aged 50 years or older with doctor-diagnosed hypertension and known medication status at follow-up were included. Analytic samples reached 4,532 participants in CHARLS, 26,128 in SHARE, and 9,495 in HRS. Non-persistence at follow-up was common in every cohort, ranging from 12.6% in HRS to 22.8% in CHARLS.
A single sentence captures the core result: medication behavior predicted future medication behavior better than baseline phenotype did.
Simple History Signals Carried More Predictive Weight Than 21 Cross-Sectional Features
Investigators compared multiple models built from different feature sets. Cross-sectional models used 21 harmonized baseline features, while history-based models incorporated 11 longitudinal variables, including the prior non-use ratio. A parsimonious three-feature score distilled the prediction to baseline non-use, prior non-use ratio, and prior discontinuation. Logistic regression and random forest models were evaluated using nested cross-validation, and one CHARLS-trained model was transported directly to SHARE and HRS without retraining.
Within each cohort, adding medication history sharply improved discrimination. AUC values rose from 0.59–0.67 for cross-sectional models to 0.78–0.85 for history-enhanced models, a gain of 0.17–0.20 with all 95% confidence intervals above zero. Those are not incremental gains from model tuning; they are the kind of jumps that tell clinicians where the real signal lives. Prior medication behavior is a direct proxy for adherence, access, tolerability, health beliefs, refill friction, and the practical burden of chronic therapy. In hypertension care, those factors often dominate over static baseline measurements.
The baseline carry-forward rule alone achieved AUCs of 0.70–0.80. The three-feature score performed even better, with AUCs of 0.75–0.83, and it was non-inferior to the full 32-feature logistic model in all three within-cohort comparisons and in both external transport settings. The prespecified non-inferiority margin was 0.03 AUC, and the three-feature score remained within that boundary throughout. A model that simple is unusual in modern prediction research, where added variables often buy little more than complexity and fragility.
Transport to Other Health Systems Favored History-Based Models
External transport separated the robust predictors from the brittle ones. When the CHARLS-trained models were applied to SHARE and HRS without retraining, the history-based models retained useful discrimination, with AUCs of 0.73–0.76 in SHARE and 0.83–0.85 in HRS. Cross-sectional models, by contrast, collapsed to near chance in SHARE, with AUCs of 0.55–0.57. In practical terms, clinical profiles that looked useful within one setting failed when moved across health systems, while the medication history core remained informative.
The reason is clinically intuitive. Cross-sectional features such as comorbidity burden, age, and other baseline descriptors can differ in meaning across countries, care systems, and survey structures. Prior medication use and discontinuation, however, directly encode the patient’s interaction with treatment over time. A history of not taking antihypertensive medication is not a passive demographic trait; it is an observed record of behavior, access, and persistence under real-world conditions. That makes it far more transportable than a static profile assembled from broad clinical descriptors.
European Country-Level Variation Exposed the Limits of Transfer Without Recalibration
The analysis did not stop at pooled European data. Across 18 SHARE countries, country-level validation showed a pooled AUC of 0.55 for cross-sectional models with I² of 30%, while history-based models pooled at 0.74–0.76 with I² of 82–88%. The high heterogeneity in the history-based models signals that even strong predictors can behave unevenly once local systems, prescribing habits, and patient behavior enter the picture. Predictive discrimination was strongest in higher-resource settings, and it worsened as non-persistence burden increased.
Two correlations were especially relevant. Transported AUC rose with national wealth, with r=0.58 and P=0.011. Calibration slope fell as non-persistence burden increased, with r=-0.69 and P=0.002, and also declined as the share of out-of-pocket expenditure increased, with r=-0.51 and P=0.032. In other words, the model could still rank risk reasonably well in wealthier systems, but the predicted probabilities became less trustworthy where stopping therapy was more common and where patients carried more direct cost burden. In a clinic, that distinction matters because a model that sorts patients correctly but misstates absolute risk can mislead decisions about follow-up intensity or outreach.
Why the Three-Variable Score May Be Enough at the Bedside
The most clinically useful message is not that machine learning failed. It is that machine learning added little beyond a tightly defined medication-history rule. The three-feature score matched the full logistic model closely enough to satisfy non-inferiority, and the simplest baseline carry-forward rule already performed surprisingly well. For clinicians and health systems, this means the first question may not be “Which algorithm should we deploy?” but “Which medication history fields do we already collect reliably, and can we act on them now?”
This has obvious implications for hypertension management workflows, particularly in settings where data infrastructure is limited. A compact history-based rule can be embedded in registries, primary care outreach lists, or pharmacy follow-up systems without needing a broad laboratory or phenotypic panel. The biological and behavioral explanation is straightforward: persistence is a longitudinal property, and longitudinal properties tend to repeat. Prior discontinuation is one of the strongest available indicators of future discontinuation because it captures habit, access, perceived benefit, side effects, and the competing demands that shape chronic self-management.
Still, the preprint leaves an important caution in place. It is a preprint, not a peer-reviewed trial, and the strongest-performing models still needed local recalibration in high-burden, lower-resource settings before individual-level deployment. The data show clear transportability of medication history, but they also show that prediction probabilities drift when health-system context changes. That leaves implementation work, not model complexity, as the next obstacle for hypertension persistence tools.