Privacy-preserving Sequential Pattern Mining in distributed EHRs for Predicting Cardiovascular Disease.

Eric W. Lee, Li Xiong, Vicki Hertzberg, Roy L. Simpson, Joyce C. HoPubMed 2021

TL;DR

This paper introduces a novel framework for privacy-preserving sequential pattern mining (SPM) on distributed Electronic Health Records (EHRs) to predict cardiovascular disease (CVD). By extracting and aggregating discriminative patterns across multiple data sources while applying differential privacy, the framework achieves nearly the same predictive accuracy as a centralized approach, significantly enhancing patient privacy.

What Problem Does This Paper Address?

Healthcare research heavily relies on Electronic Health Records (EHRs) to discover relationships between patient conditions, treatments, and outcomes, particularly for tasks like risk prediction. However, a major hurdle is that EHR data is often fragmented across multiple institutions or data warehouses, making it challenging to mine comprehensive patterns. Compounding this issue are stringent privacy regulations that prevent direct sharing of raw patient data. Existing SPM methods typically assume a centralized data repository, which is impractical in real-world distributed healthcare systems. This paper addresses the critical need for a method that can effectively mine distributed EHRs for predictive modeling while rigorously protecting patient privacy.

Methods

The authors propose a two-pronged approach:

  1. SPM-based Framework with No Aggregation: This initial approach trains risk prediction models independently at each data source. Discriminative patterns are extracted locally ("local discriminative patterns"), and the top k patterns from each source are used as features for a local classifier. Crucially, no pattern information is shared between sources, ensuring maximum privacy but potentially missing valuable insights from other datasets.
  2. Privacy-preserving SPM-based Framework with Aggregation: This is the core contribution. It extends the first approach by introducing a centralized aggregation step.
  3. Local Pattern Extraction: Similar to the first approach, SPM is applied to each data source to extract frequent subsequent patterns and then local discriminative patterns.
  4. Differential Privacy (DP): To protect privacy, DP mechanisms (Laplace, Exponential, and SVT) are applied to the support counts of these patterns before they are shared. This means that while patterns are aggregated, the exact frequency from any single source is obscured by added noise.
  5. Global Pattern Aggregation: The noisy local discriminative patterns are sent to a centralized server. Here, patterns from all sources are aggregated, and the top k global discriminative patterns are selected.
  6. Federated Learning for Prediction: These global discriminative patterns are then shared back to each data source to train local classifiers. This allows each source to benefit from patterns discovered across the entire distributed system while maintaining privacy.

The paper also details the use of the FuzzyGap sequence representation for handling temporal data and the use of Clinical Classifications Software (CCS) codes for better interpretability of diagnosis codes.

Datasets

The study utilizes the Project NeLLTM dataset, which contains de-identified electronic health records from over 1 million patients at Emory Healthcare spanning from 2012 to 2018. This dataset includes structured (lab values) and unstructured (clinical notes, radiology reports) text. For the case study, patients with type 2 diabetes were identified, and a case-control cohort was constructed to predict the development of cardiovascular disease (CVD). A 4:1 matching ratio of non-CVD to CVD patients was used to address class imbalance. The dataset was partitioned into 2, 4, and 8 sources for evaluation.

Key Insights and Ideas

  • Privacy-Preserving Aggregation is Key: The most significant insight is that aggregating noisy discriminative patterns (using DP on support counts) can achieve nearly the same predictive performance as a non-private, centralized approach, while significantly enhancing privacy.
  • Global Patterns Outperform Local: The "no aggregation" framework, while private, suffers from a substantial drop in predictive power as the number of partitions increases. This highlights the critical importance of leveraging discriminative patterns discovered across all data sources for robust prediction.
  • DP Trade-off is Manageable: Applying DP mechanisms does introduce a slight decrease in predictive performance compared to a non-private aggregated approach. However, this trade-off is marginal and significantly better than the performance degradation seen in the "no aggregation" scenario.
  • Pattern Recoverability Matters: The study emphasizes the importance of recovering discriminative patterns. DP mechanisms, while protecting privacy, can affect the number and quality of patterns recovered, impacting the overall predictive performance.

Key Results

  • Predictive Performance: The privacy-preserving aggregated framework (with DP) achieved predictive performance (F1 score and AUC) that was very close to the non-private aggregated framework. For instance, in the 8-partition setting, the F1 scores were comparable, and AUC showed only a marginal decrease.
  • Impact of Partitions: As the number of partitions increased, the "no aggregation" framework saw a drastic decline in predictive power. In contrast, the aggregated frameworks (with and without privacy) showed a much more stable performance, with privacy-preserving versions exhibiting only a slight, manageable decrease.
  • Pattern Recoverability: Precision and recall metrics indicated that DP mechanisms, particularly with more partitions, could reduce the number of discriminative patterns recovered. However, the aggregated approach still managed to recover a substantial number of important patterns.
  • Top k Selection: The study found that using the top 100 discriminative patterns from each class (total 200 features) provided a good balance between predictive power and avoiding overfitting, with further increases in k yielding diminishing returns.

Limitations and Future Work

  • Single Real-World EHR Source: The study used data from a single institution (Emory Healthcare). The authors acknowledge that real-world EHRs are often more heterogeneous, and using data from diverse institutions would provide a more robust evaluation.
  • Even Partition Sizes: The experiments used evenly sized partitions. In practice, data sources can vary significantly in size, which could lead to local patterns from larger sources dominating global patterns. The authors suggest exploring weight-based aggregation to mitigate this.
  • ICD-9 Codes Only: The study primarily used ICD-9 codes. Incorporating other information like procedure codes or prescriptions could enhance predictive power but would require more computational resources.
  • Future Work: The authors suggest exploring more sophisticated aggregation strategies, incorporating a wider range of patient data (procedures, prescriptions), and evaluating the framework on more heterogeneous, multi-institutional datasets.

Summary updated

Abstract

From electronic health records (EHRs), the relationship between patients' conditions, treatments, and outcomes can be discovered and used in various healthcare research tasks such as risk prediction. In practice, EHRs can be stored in one or more data warehouses, and mining from distributed data sources becomes challenging. Another challenge arises from privacy laws because patient data cannot be used without some patient privacy guarantees. Thus, in this paper, we propose a privacy-preserving framework using sequential pattern mining in distributed data sources. Our framework extracts patterns from each source and shares patterns with other sources to discover discriminative and representative patterns that can be used for risk prediction while preserving privacy. We demonstrate our framework using a case study of predicting Cardiovascular Disease in patients with type 2 diabetes and show the effectiveness of our framework with several sources and by applying differential privacy mechanisms.
BibTeX
@inproceedings{lee2021privacy,
  author = {Eric W. Lee and Li Xiong and Vicki Hertzberg and Roy L. Simpson and Joyce C. Ho},
  title = {{Privacy-preserving Sequential Pattern Mining in distributed EHRs for Predicting Cardiovascular Disease.}},
  booktitle = {PubMed},
  year = {2021},
}
Figure