Undiagnosed hypertension screening drops missed cases from 1,237 to 729 with positive-unlabeled learning

The screening model no longer treats everyone without a prior diagnosis as non-hypertensive. When labels are built from prior diagnosis alone, undiagnosed hypertensive participants enter training as negatives, so the classifier learns something closer to whether a person was ever diagnosed than whether that person has hypertension. In an independent test set of 8,611 participants, the pipeline reached a sensitivity of 0.817, above logistic regression at 0.683, standard XGBoost at 0.690, and a neural network at 0.719.

The training sample is rebuilt to keep only high-confidence candidate positives and candidate negatives. A single-stage procedure assigns pseudo-labels to every unlabeled observation, forcing samples near the decision boundary into one class and passing pseudo-label error to the downstream classifier. The two-stage reconstruction retained 8,017 pseudo-positive and 8,017 pseudo-negative observations, 16,034 participants or 61.9% of the original training set, while 9,849 highly ambiguous observations were excluded from the final supervised loss.

A linear model ranks globally first, and a nonlinear model then handles only the uncertain middle region. Elastic Net logistic regression produces a positive-versus-unlabeled score and splits the training sample at thresholds of 0.076 and 0.598 into candidate positives, candidate negatives, and an uncertain set. A random forest re-ranks only that uncertain set of 18,426 participants, and thresholds near 0.242 and 0.460 yield 2,704 stage-2 candidate positives, 5,873 candidate negatives, and 9,849 rejected observations before XGBoost is trained on the balanced pseudo-labeled set.

The pipeline targets community-dwelling adults aged 45 years and older from eight study areas in China, with inputs limited to demographic, lifestyle, medical-history, and anthropometric variables; laboratory measurements, medication information, and blood-pressure measurements are excluded from the predictors. It is positioned as risk stratification before confirmatory blood-pressure measurement rather than a substitute for clinical diagnosis. The authors propose evaluating sensitivity, false-negative reduction, and specificity in independent external datasets, comparing the framework systematically with alternative positive-unlabeled, weak-supervision, and mixture-model approaches, and optimizing decision thresholds against the relative costs of missed cases, confirmatory measurement, and acceptable false-positive burden.

Positive-unlabeled learning assumes labeled positives are reliable enough, yet whether a hypertensive person has been previously diagnosed is unlikely to be random; the labeling mechanism may vary with health-care access, health awareness, geography, and socioeconomic status, producing positive-selection bias. Two-stage sample selection, rejection of uncertain observations, and 1:1 class balancing jointly alter the training-set composition and class prior, so the sensitivity gain cannot be attributed uniquely to any single component without component-specific ablation analyses. The lower specificity may increase false-positive volume and confirmatory workload at large scale, and net practical value depends on the screening cost structure. The study population comprised community-dwelling adults aged 45 years and older in China, so findings should not be extrapolated directly to younger populations or other settings.