APH-Net: a multi-task uncertainty-weighted neural architecture for subgroup-level pediatric health surveillance modeling
Year: 2026
Type: article
Abstract: Aggregated public health tables contain repeated demographic and temporal estimates, but random record-level splits can overstate predictive performance when the same subgroup definitions occur in both training and test data. This study evaluates APH-Net, a condition-conditioned multi-task neural architecture, using a CDC surveillance table with 1,203 condition-stratum-period cells across eight health conditions, 25 marginal subgroup strata, and seven reporting windows. The primary evaluation withheld five complete subgroup strata (197 cells), while model selection used only the remaining 1,006 cells. XGBoost achieved the strongest primary regression performance (MAE = 1.272, RMSE = 1.699, R² = 0.862; 95% CI 0.823–0.887 for R²), followed by APH-Net (R² = 0.853) and Random Forest (R² = 0.838). For the auxiliary above-training-median classification task, XGBoost achieved ROC-AUC = 0.775 and macro-F1 = 0.626, whereas APH-Net achieved ROC-AUC = 0.627. Repeated subgroup holdouts and a non-overlapping temporal test produced lower and more variable performance than random-split estimates. APH-Net ablation showed task-dependent value: classification discrimination improved over a single-head MLP, while regression did not. The findings support subgroup-level surveillance modeling but not individual clinical risk prediction.
Source: Scientific Reports
Authors: Liying Wang, Jihe Kang, Quantao He
Institutions: Guangdong Polytechnic Normal University, Shenzhen University
Language: English
Subfield: Epidemiology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7230943897
