Machine Unlearning Based on Globally Refined Convergent Clustering for Health Survey Data-driven Prediction Model

Machine learning models built from national health surveys enable population-scale risk stratification, yet the GDPR's "right to be forgotten" mandates that participants' data and their downstream impacts be fully erasable upon request. Existing privacy-preserving and unlearning approaches remain limited because they can be computationally prohibitive, provide only approximate protection by obscuring individual influence, rely on parameter-level approximation, or depend on localized retraining whose effectiveness is sensitive to data partitioning. In health survey settings, cohort heterogeneity across sociodemographic, behavioral, and clinical dimensions can further induce imbalanced shard distributions, thereby compromising the stability and effectiveness of unlearning. To overcome these challenges, we propose a subgroup-aware exact unlearning framework for heterogeneous health survey data, in which GRC uses a multi-stage refinement process to discover epidemiologically coherent subgroups and construct proportion-preserving shards for localized unlearning. These subgroup labels facilitate proportional shard partitioning, effectively preserving cohort ratios and ensuring balanced representation within each shard. Each shard undergoes incremental training with intermediate model snapshots periodically recorded; upon receiving a deletion request, only affected shards roll back to their nearest clean snapshot and undergo localized retraining, followed by aggregation via a soft-voting ensemble to restore global consistency. This design provides a structured exact unlearning workflow with bounded retraining scope and empirically strong forgetting performance while maintaining predictive performance. Evaluations on depression-risk prediction tasks derived from the NHANES and CHARLS datasets demonstrate that our approach consistently improves Zero-Retrain-Forgetting performance while preserving predictive utility, supporting an effective and practical exact unlearning strategy for prediction models driven by health survey data.