DP-NegGrad+: Differentially Private Machine Unlearning with a Post-Quantization Robustness Check for Diabetes Risk Prediction

Once a patient’s data has been used to train a clinical risk-prediction model, it is generally difficult to guarantee that the model has truly forgotten that patient if they later withdraw consent. Machine unlearning addresses this problem, but most evaluations of unlearning quality stop at the full-precision model, even though edge and point-of-care devices routinely quantize a model before it is shipped. In this work, we apply cohort-level machine unlearning to a diabetes risk-prediction model, where the forget cohort is patients whose smoking history is recorded as “former”. Beyond the standard Original, Retrain, Fine-Tune, and NegGrad+ comparison used in the unlearning literature, we add two components relevant to real clinical deployment: DP-NegGrad+, a DP-SGD-inspired noisy-gradient variant of NegGrad+ that adds a privacy margin for the retained cohort and not only the forgotten one, and a post-unlearning INT8 quantization robustness check, since edge and point-of-care deployment routinely quantizes models before shipping them. Across five settings (Original, Retrain, Fine-Tune, NegGrad+, DP-NegGrad+) we report test accuracy, F1, forget-set accuracy, and membership-inference AUC, both before and after INT8 quantization, and we verify explanation fidelity with SHAP. Experimental results demonstrate that DP-NegGrad+, our proposed method, tracks NegGrad+ closely while attaining 0.8796 test accuracy with a membership-inference AUC of 0.5357, adding a calibrated Gaussian noise mechanism (noise multiplier σ = 1.0, δ = 10⁻⁵) that gives the retained cohort an explicit, quantifiable privacy margin NegGrad+ does not provide, without sacrificing utility, forgetting quality, or explanation fidelity, even after INT8 quantization.