How Much Does Privacy Cost? An Empirical Study of Client-Level Differential Privacy and Quantization in Cross-Silo Federated Clinical Risk Prediction

Federated learning (FL) is often proposed as the natural way to train clinical risk-prediction models across hospitals that cannot pool patient data, but adding formal differential privacy (DP) and communication-efficient quantization on top of FL is usually evaluated only on a single, often synthetic, dataset. We present an empirical, privacy-accounted study of the utility, privacy, and communication trade-offs of FedAvg, FedProx, client-level DP-FedProx, and 8-bit quantized DP-FedProx (Q-DP-FedProx) across three tabular clinical risk-prediction tasks: two real multi-site cross-silo datasets from the FLamby suite (Fed-Heart-Disease, 4 sites; Fed-TCGA-BRCA, 6 sites) and one externally partitioned dataset (Wisconsin Diagnostic Breast Cancer) used for a controlled non-IID robustness sweep. Using Rényi DP accounting with no subsampling amplification, we find that a fixed noise multiplier that costs only 1.6 points of AUC on the large, clean, synthetically partitioned dataset costs 11.4 points of AUC on the smallest real multi-site dataset (Fed-Heart-Disease, n=486 across 4 sites, one site with only 30 patients), at matched ε ≈ 2.2, δ=10⁻⁵. Quantization contributes negligible additional utility loss while cutting uplink communication by 2.5–3.3×. We release a fully reproducible pipeline, including a closed-form per-patient local explanation method for the linear federated models and an explicit audit of which demographic covariates (including self-reported race) drive individual risk predictions in the Fed-TCGA-BRCA model.