Predicting Nocturnal Hypertension in CKD Through Table-Value Diffusion Model

BACKGROUND: Ambulatory blood pressure monitoring is indispensable for diagnosing nocturnal hypertension (NH) among patients with chronic kidney disease, but it is costly and time-consuming. Screening tools for predicting high-risk NH are urgently needed. METHODS: A large cohort of 5769 patients with nondialysis chronic kidney disease were enrolled (hospital A: 4565; hospital B: 1204). Patients from hospital A were split 8:2 into training/internal test sets; hospital B served as the external test set. A total of 1006 patients with at least 2 valid ambulatory blood pressure monitoring recordings were used for secondary validation. We proposed a table-value diffusion model by incorporating generative modeling concepts to map distributions between clinical variables and predict NH. The predicted probabilities for NH were stratified and validated according to patients’ true adverse renal/cardiovascular prognoses. RESULTS: Through a rigorous selection process combining univariate ranking and clinical expertise, we selected 7 core predictors: age, body mass index, clinic systolic/diastolic blood pressure, estimated glomerular filtration rate, hypertension history, and use of non–renin-angiotensin-aldosterone system antihypertensive agents. In the test sets, the 7-variable diffusion model achieved a significantly higher area under the curve than the 7-variable logistic model (internal: 0.870 versus 0.807, P