[論文レビュー] Federated and Differentially Private Learning for Electronic Health Records

Stephen Pfohl, Andrew M. Dai+1 more|arXiv (Cornell University)|Nov 13, 2019

Privacy-Preserving Technologies in Data参考文献 10被引用数 44

ひとこと要約

本論文は eICU-CRD データセットを用いた臨床予測タスクにおける中央集権型、連合学習、および差分プライバシーを用いた訓練を比較し、DP-SGD は中央集権的には有効だが、個人データ保護を前提とする学習において連合設定では課題があることを示している。

ABSTRACT

The use of collaborative and decentralized machine learning techniques such as federated learning have the potential to enable the development and deployment of clinical risk predictions models in low-resource settings without requiring sensitive data be shared or stored in a central repository. This process necessitates communication of model weights or updates between collaborating entities, but it is unclear to what extent patient privacy is compromised as a result. To gain insight into this question, we study the efficacy of centralized versus federated learning in both private and non-private settings. The clinical prediction tasks we consider are the prediction of prolonged length of stay and in-hospital mortality across thirty one hospitals in the eICU Collaborative Research Database. We find that while it is straightforward to apply differentially private stochastic gradient descent to achieve strong privacy bounds when training in a centralized setting, it is considerably more difficult to do so in the federated setting.

研究の動機と目的

• 病院間で生データを共有せずに予測モデリングを行うための連携を動機づけつつ、プライバシー保証の課題にも対処する。
• 非プライベートおよび DP 設定における中央集権型 vs. 連合学習を臨床タスクで評価する。
• 中央集権型および連合設定における差分プライバシー適用時のプライバシーと有用性のトレードオフを定量化する。

提案手法

• Utilize federated averaging and DP-SGD to train logistic regression and one-hidden-layer networks on 31 hospitals from eICU-CRD.
• Construct features from lab orders/results, medications, diagnoses, and patient demographics within 24 hours of admission.
• Compare local, centralized, and federated training, with and without differential privacy, across two tasks: in-hospital mortality and prolonged length of stay.
• Assess privacy via DP parameters (epsilon, delta) and model performance via AUC-ROC with DeLong confidence intervals.

実験結果

リサーチクエスチョン

• RQ1 Can federated learning provide similar predictive performance to centralized learning on ICU outcome prediction tasks?
• RQ2 How does the addition of differential privacy affect model utility in centralized vs. federated settings?
• RQ3 Is DP-SGD practical in federated learning for clinical data in terms of privacy guarantees and accuracy?

主な発見

• Federated learning often yields AUC-ROC close to centralized learning and frequently improves over local hospital models, though gains are not always statistically significant.
• Differentially private centralized training can achieve strong privacy (epsilon around 1) with modest AUC-ROC degradation (e.g., stay: 0.763 vs 0.73; mortality: 0.876 vs 0.832).
• Federated learning with DP-SGD often performs poorly in both AUC-ROC and privacy cost (epsilon) for mortality prediction compared to non-DP settings.
• DP-SGD hyperparameters optimized on local data do not readily transfer to federated settings, highlighting challenges in privacy-utility trade-offs for distributed clinical data.
• The study emphasizes that ignoring the privacy cost of model selection can yield optimistic privacy estimates.