Clinical History Improves Diabetes Hypertension Risk Prediction

Key Takeaways

• Among adults with type 2 diabetes at a single Chinese center, historical measurements improved internally validated prediction most clearly for first recorded hypertension.
• MASLD showed a smaller pooled gain from history, but its later landmark had few events.
• Coronary heart disease, diabetic kidney disease, and diabetic retinopathy showed no comparable pooled improvement from history.

Fu and colleagues' diabetes landmarking study drew on a retrospective cohort at a tertiary hospital in Shenzhen, China: 646 adults aged at least 40 years with type 2 diabetes and complete annual records from baseline through year 3. At successive annual landmarks, investigators predicted the first clinician-recorded positive status at the next annual visit for hypertension, coronary heart disease, diabetic kidney disease, diabetic retinopathy, or metabolic dysfunction-associated steatotic liver disease (MASLD). Each outcome had its own risk set, excluding people with a prior positive record of that condition. Pooled person-landmark records could include the same patient more than once.

At the year-1 and year-2 landmarks, investigators built progressively richer predictor sets: landmark-only information; fixed baseline characteristics of sex, age, diabetes duration, and height; current-visit measurements and non-target comorbidity information; and a history set adding earlier means and recent changes. Elastic-net logistic regression and Light Gradient Boosting Machine (LightGBM) estimated outcome-specific risk over the next annual interval.

Patient-grouped five-fold outer cross-validation, repeated 10 times, separated test patients from model development; inner validation guided tuning. Within each algorithm, the history-minus-current comparison aligned outcome-specific risk sets, outer test patients, preprocessing, tuning, calibration, and evaluation records. Separate checks comprised fully nested configuration selection inside outer training data and 500 patient-level development bootstrap replicates evaluated in out-of-bag patients.

For first recorded hypertension, earlier measurements produced the clearest paired gain in area under the receiver operating characteristic curve (AUC). With elastic-net, the pooled history-minus-current AUC difference was 0.097, with a 95% confidence interval (CI) of 0.057–0.136; with LightGBM, it was 0.118 (95% CI 0.081–0.158).

For MASLD, the pooled history-minus-current AUC difference was 0.022 (95% CI 0.002–0.042) for elastic-net and 0.025 (95% CI 0.009–0.041) for LightGBM. The later MASLD landmark contained 15 events, limiting what its individual comparison could establish.

Coronary heart disease, diabetic kidney disease, and diabetic retinopathy had no comparable pooled history gain. LightGBM had a modest AUC advantage for coronary heart disease and MASLD; elastic-net had higher average precision for retinopathy. Algorithm performance was similar for hypertension and kidney disease. Selected representative models' pooled out-of-fold discrimination and near-unity calibration described internal, post-selection performance. Separately, fully nested selection most often favored history-LightGBM for hypertension and MASLD but varied more for coronary disease. Out-of-bag development bootstraps favored history more consistently for hypertension than for MASLD.

Complete follow-up at one Shenzhen center narrowed the population assessed: year-1 eligibility depended on later year-3 availability, although year-3 measurements were not year-1 predictors. Annual first-positive electronic health record (EHR) fields marked recorded status, not independently adjudicated onset. Adding measures close to a contemporary diagnosis sharply raised apparent coronary and kidney-disease discrimination. Selected-model absolute metrics were chosen partly using the same out-of-fold data; intervals calculated from completed predictions did not repeat model development, and comparisons lacked formal multiplicity adjustment. Decision curves lacked specified actions or validated thresholds, and improved discrimination did not demonstrate patient benefit.

Fu and colleagues identified independent temporal and multicenter validation, followed by prospective assessment of clinical impact, as remaining steps.

Clinician Questions

Did year-3 measurements leak into the year-1 diabetes risk predictions?

In a leakage audit, investigators deliberately changed year-3 dynamic values and rebuilt the year-1 datasets. Patient membership, outcome labels, predictor columns, and predictor values remained unchanged. The audit did not address selection related to requiring year-3 follow-up for cohort inclusion.

How were missing clinical measurements handled in the diabetes prediction models?

Elastic-net used median imputation, missingness indicators, and scaling learned within training folds; LightGBM handled missing values natively. Future-visit values and cross-year interpolation were not used to fill missing earlier measurements. These procedures concerned predictors, not adjudication of missing outcomes.

Which historical measurements contributed to hypertension and MASLD prediction in type 2 diabetes?

Prior triglycerides and uric acid and a recent creatinine change ranked prominently for hypertension. Earlier and current body mass index and a recent total-cholesterol change were prominent for MASLD. These were model-specific predictive importance patterns, not independently established causal effects.