DP-RuL: Differentially-Private Rule Learning for Clinical Decision Support Systems

Josephine Lamp,Lu Feng,David Evans

2024-05-16

Abstract:Serious privacy concerns arise with the use of patient data in rule-based clinical decision support systems (CDSS). The goal of a privacy-preserving CDSS is to learn a population ruleset from individual clients' local rulesets, while protecting the potentially sensitive information contained in the rulesets. We present the first work focused on this problem and develop a framework for learning population rulesets with local differential privacy (LDP), suitable for use within a distributed CDSS and other distributed settings. Our rule discovery protocol uses a Monte-Carlo Tree Search (MCTS) method integrated with LDP to search a rule grammar in a structured way and find rule structures clients are likely to have. Randomized response queries are sent to clients to determine promising paths to search within the rule grammar. In addition, we introduce an adaptive budget allocation method which dynamically determines how much privacy loss budget to use at each query, resulting in better privacy-utility trade-offs. We evaluate our approach using three clinical datasets and find that we are able to learn population rulesets with high coverage (breadth of rules) and clinical utility even at low privacy loss budgets.

Cryptography and Security

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: when using patient data in Clinical Decision Support Systems (CDSS), how to learn a group rule set with wide coverage and clinical practicality from the local rule sets of multiple clients under the premise of protecting individual privacy. Specifically, the paper mentions: 1. **Privacy issues**: Since patient data contains sensitive information, direct use of this data may lead to privacy leakage. Especially in third - party health applications, the privacy risk is higher as they are not protected by the Health Insurance Portability and Accountability Act (HIPAA). 2. **Objective**: The research aims to develop a framework. In a distributed CDSS environment, through Local Differential Privacy (LDP) technology, the server can collect and learn useful group rule sets from multiple clients without accessing individual sensitive information. 3. **Method**: The authors propose a method that combines Monte Carlo Tree Search (MCTS) and LDP. It discovers potential rule paths through random response queries and introduces an adaptive privacy loss budget allocation method to optimize the trade - off between privacy and utility. 4. **Evaluation**: The researchers conducted experiments using clinical datasets in three different fields (intensive care unit, sepsis, type 1 diabetes). The results show that even with a lower privacy loss budget, this method can still learn group rule sets with high coverage and clinical practicality. Therefore, the core problem of this paper is how to effectively use distributed patient data to improve the performance of clinical decision support systems while ensuring privacy protection.

DP-RuL: Differentially-Private Rule Learning for Clinical Decision Support Systems

DPMLBench: Holistic Evaluation of Differentially Private Machine Learning

Chasing Your Long Tails: Differentially Private Prediction in Health Care Settings

Locally Differentially Private Distributed Deep Learning via Knowledge Distillation

Differentially Private Reinforcement Learning with Self-Play

Scalable Differential Privacy Mechanisms for Real-Time Machine Learning Applications

Inference With Combining Rules From Multiple Differentially Private Synthetic Datasets

Offline Reinforcement Learning with Differential Privacy

(Local) Differential Privacy has NO Disparate Impact on Fairness

DPDR: Gradient Decomposition and Reconstruction for Differentially Private Deep Learning

Local Differential Privacy for data collection and analysis

A Differentially Private Weighted Empirical Risk Minimization Procedure and its Application to Outcome Weighted Learning

Differentially Private Learning with Small Public Data.

LazyDP: Co-Designing Algorithm-Software for Scalable Training of Differentially Private Recommendation Models

Stochastic Differentially Private and Fair Learning

DPCL: Contrastive Representation Learning with Differential Privacy

Locally Private Distributed Reinforcement Learning

Differentially Private Reinforcement Learning with Linear Function Approximation

A Systematic and Formal Study of the Impact of Local Differential Privacy on Fairness: Preliminary Results

Differentially private outcome‐weighted learning for optimal dynamic treatment regime estimation

LLM-based Privacy Data Augmentation Guided by Knowledge Distillation with a Distribution Tutor for Medical Text Classification