20K Distributed Learning Challenge

All current innovations in medicine, including personalized medicine; artificial intelligence; (Big) data driven medicine; learning health care system; value based health care and decision support systems, rely on the sharing of data across health care providers. But sharing of data is hampered by administrative, political, ethical and technical barriers(Sullivan et al., 2011). This limits the amount of health data available for the above innovations and life sciences in general as well as other secondary uses such as quality improvement.

The investigators hypothesize that sharing questions rather than sharing data is a better approach and can unlock orders of magnitude more data while limiting privacy and other concerns. An infrastructure to bring questions to the data has been demonstrated to work recently in project such as euroCAT(Lambin et al., 2013; Deist et al., 2017), Datashield (Gaye et al., 2014) and OHDSI (Hripcsak et al., 2015). However, the scale of the prior work has been limited in terms of the number of data subjects, number of data providers and global coverage.

In the experience of the investigators, the main challenges of scaling up the infrastructure are 1) the effort necessary to make data FAIR at each site ("stations"), 2) the technical and legal governance ("track") and 3) the mathematics and engineering of learning applications ("trains") - together called the Personal Health Train (PHT) infrastructure. Since multiple years a global consortium of healthcare providers, scientists and commercial parties called CORAL (Community in Oncology for RApid Learning) have worked on all three PHT challenges.

The aim of this study is to show that the PHT distributed learning infrastructure can be scaled to many 1000s of patients, specifically the investigators aim to machine learn a predictive model from more than 20.000 non-small cell lung cancer patients from more than 5 health care providers from more than 5 countries.