1 Objectives
This document describes the Privacy-preserving Distributed Algorithms (PDA) framework and how the end-users should use it for conducting a federated learning analysis with distributed multi-center data. The PDA website is at https://pdamethods.org/, the software package pda is at https://github.com/Penncil/pda, and the secure data transferring platform is at https://pda-ota.pdamethods.org/.
1.1 Terminologies
Table 1 contains the terms and acronyms that are used throughout this guideline.
| Terms | Meaning |
|---|---|
| RWD | real-world data |
| CRN | clinical research network |
| DUA | data use agreement |
| PDA | Privacy-preserving distributed algorithm |
| Site | The minimum unit of which data privacy should be preserved, e.g. institution, hospital, patient cohort, or database etc. PDA assumes a lead site and multiple collaborative sites, where the lead site coordinates the data communication and federated analysis. |
| IPD | Individual-participant data, of which the transferring across sites are prohibited by privacy regulations such as HIPPA and GDPR |
| AD | Aggregate data, as opposed to IPD, AD at site level protects the privacy of the IPD. |
| pooled | Pooled data analysis means the multi-site IPD data can be combined together to conduct analysis |
| One-shot | One (or a few) round(s) of data communication. Not confused with the One-shot learning (OSL) as a classification technique in machine learning |
| meta | Meta-analysis approach to conduct analysis using IPD at each site separately and then take average of the results. |
| OTA | Over-the-air, the platform for data transferring when conducting a federated learning using PDA method. |
1.2 Organization of this guideline
Section 2 will provide an Overview of the PDA Framework including the Background of RWD, CRN and federated learning, Core advantages and impact of PDA. Figure 1 below is a diagram showing a general project pipeline of using PDA to conduct a federated-learning for your study. The steps will be elaborated in the next sections. Specifically, Section 3 will provide an overview of the PDA framework and how to decide the appropriate PDA model. The extensive real-world application examples will help build the PDA study protocol. Section 4 introduces the software development of the ‘pda’ package with extensive code demo examples. Section 5 introduces the usage of the PDA-OTA platform for a seamless federated learning combining the analysis and data transferring. More of the technical details can be found in the Section 6 References, and Section 7.1. The templates of PDA protocol and legal documents are also provided in Section 7.
Figure 1. A general project pipeline of conducting a federated learning analysis for your study using the PDA framework.
[Note: The PDA framework requires the data from multiple sites to be harmonized to have the same format (e.g. variable names, category coding) before conducting analysis. Data harmonization were usually done with a common data model (CDM), such as xxx. We recommend using the Observational Medical Outcomes Partnership (OMOP) CDM and the associated tools of the Observational Health Data Sciences and Informatics (OHDSI) initiative due to its open-source and continuous development. Although data harmonization is crucial for data integration, we deem it as upstream work and don’t focus on it in this guideline.]