PRISM-VFL: A Differentially Private Vertical Federated Framework for Heterogeneous Multi-Task Clinical Prediction
Abstract
Abstract Vertical federated learning (VFL) enables collaborative clinical model training across institutions holding complementary patient features. However, existing frameworks lack support for the concurrent optimisation of heterogeneous prediction tasks under formal privacy guarantees, limiting their applicability in multi-site deployments. This study introduces PRISM-VFL (Privacy-Resilient Institutional Split Multi-task VFL), a VFL framework that pairs per-site long short-term memory (LSTM) encoders with a server-side Multi-gate Mixture-of-Experts (MMoE) module and applies task-stratified differentially private stochastic gradient descent (DP-SGD) at the cut layer. The framework was evaluated on vertically partitioned MIMIC-III data across three simulated hospitals with distinct outcomes, namely in-hospital mortality (IHM), Decompensation risk stratification, Phenotyping, alongside a theoretically derived multi-task label inference bound, g(σ, ρ) = Φ(C √ (1 + ρ)/σ). PRISM achieved mean area under the receiver operating characteristic curve AUC-ROCs of 0.781 for IHM,0.632 ± 0.116 for Decompensation (high cross-seed variance); and macro-averaged AUC-ROC (macro-AUC) 0.613 for Phenotyping. Task-specific utility analysis demonstrated that IHM maintained clinical viability under moderate privacy thresholds (ε∗ ∈ (1, 2)), while Decompensation and Phenotyping did not reach utility floors even in the absence of differential privacy (DP) constraints. Under strong privacy (ε =…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.