2 Overview of the PDA framework

[Note: This section is for general users]

2.1 Background: federated learning

The recent decade has seen an increasing trend in using real-world data (RWD) to generate evidence for supporting regulatory decision-making in biomedical and healthcare venue. Observational RWD often includes large amounts of patient data from a variety of sources, forming clinical research networks (CRNs). Examples are the National COVID Cohort Collaborative or N3C (https://ncats.nih.gov/n3c), the Consortium for Clinical Characterization of COVID-19 by EHR or 4CE (https://covidclinical.net/), and the Observational Health Data Sciences and Informatics (OHDSI) network (ohdsi.org). The benefits of integrating heterogeneous data sources include increasing the generalizability of results, and increasing the power to detect significant associations when studying rare outcomes or exposures.

Large-scale observational RWD studies often rely on multiple institutions to contribute data. The two main mechanisms for ‘sharing’ data include centralized data repositories and distributed data-sharing. The granularity of data shared depends on the data use agreements (DUA) among participating institutions. Some DUAs allow for all individual patient data (IPD) to be centralized, resulting in a pooling of all patient data within a cloud or at the coordinating center for the study. Other agreements prohibit sharing of IPD, instead only allow summary-level aggregate data (AD) to be shared by each institution. With the regulation of privacy protection by HIPAA or GDPR, the sharing of IPD across institutions are difficult, while sharing AD is more feasible in a distributed CRN. With this distributed data sharing mechanism, individual institutions maintain control over their own data and conduct analyses locally before sharing results with collaborating institutions.

The preferred distributed data-sharing arrangements permits the usage of federated learning methods to conduct multi-institutional analysis. In general, federated learning is a decentralized approach that allows multiple institutions or devices to train machine learning models collaboratively without sharing private data. It has been widely used in industrial scenarios such as mobile applications. Federated learning in general requires iterative transferring of the AD (e.g. the model parameters trained in local data) across sites. This often requires infrastructural development for data communication, and add extra burden for feasible federated learning in the context of a CRN.

2.2 The PDA framework: core advantages

The Privacy-preserving Distributed Algorithm (PDA, https://pdamethods.org/) framework is designed for conducting federated learning in a distributed CRN. See a video introducing the PDA framework [https://pdamethods.org/]. Compare to other alternative distributed/federated learning methods, it has 4 core advantages: it is

  1. privacy-preserving, as it transfers only aggregated data across sites;
  2. accurate, as it achieves results very close to the pooled analysis;
  3. communication-efficient, as it only requires a few (non-iteractive) rounds of data transferring; and
  4. heterogeneity-aware, as it allows the estimated model parameters to vary across sites.

Figure 1. The core advantages of the PDA framework for federated learning in a CRN.

2.3 The PDA framework: development and impact

PDA is a toolbox for conducting federated statistical/machine learning analyses between sites in a distributed CRN. With the first development in 2018, PDA now has more than 30 models (algorithms) encapsuled. The methods developed in the PDA framework have been published in top academic journals, including xxx. The methods have been applied in many applicational scenarios in epidemiological and clinical studies including outcome-exposure association analyses, disease risk prediction, hospital profiling (quality ranking), phenotype subgroup identification, disease causal inference, and clinical trial emulation using observational data, etc. These real-world applications were conducted using data from various CRNs, see the Figure 2 below for the collaborative data partners. The PDA framework was also supported by national Grants agents including xxx, and gained public impact…

Figure 2. The collaborative data partners for real-world applications of PDA methods.