DAT-ICU: Datathon for intensive care
This weeked, the ACTERRÉA research group has been engaged in DAT-ICU, a research competition organized by Assistance Publique des Hôpitaux de Paris (APHP) at the Rothschild Hospital in Paris. During this so called «datathon», teams had 38 hours to define and answer clinically relevant research questions using anonymized data from Intensive Care Units (ICU) patients.
Among the 200 participants, twenty-five researchers from five different countries joined forces in three ACTERRÉA teams. Research projects had to rely on the Medical Information Mart for Intensive Care (MIMIC) dataset made available largely through the work of researchers at the MIT Laboratory for Computational Physiology and collaborating research groups (see references below).
With this post, we would like to share with you how this exhilarating, two-day adventure unfolded. On Friday January 19th, we all gathered at a slack meeting. We introduced ourselves, Professor Romain Pirracchio briefly presented the research topics, and we agreed on a global strategy. On Saturday January 20th, eight o'clock sharp, here we were, regrouped in three teams, each consisting of clinicians and data scientists. We did not really know how we would work together, but we were enthusiastic.
First, clinicians defined the clinical problem that each team would tackle. Team 1 would focus on predicting hemodynamic instability within 14 hours after ICU admission, based on data from the first two hours. Team 2 would address the prediction of the mean arterial pressure in patients who underwent a cardiac surgery. Team 3 would validate a previously developed gravity score in patients with severe head traumas.
Second, all the members discussed and decided which patients to include in the learning set, and which variables our prediction algorithms would exploit. Reducing the dimension of the data set was crucial, to speed up the downloading time.
Third, data scientists started working on extracting a clean database meeting all the clinical requirements. This step implied constant interaction between members, to adapt clinical requirements to technical possibilities. The combination of individual initiatives and regular group meetings to share information, ask for advice or provide guidance proved very efficient eventually. Moreover, learning how to use Observational Medical Outcomes Partnership (OMOP), the APHP data format, proved long and difficult, and we ended up with little time left to implement, apply and evaluate different different machine learning algorithms. Specifically, team 1 developed an instance of the Super Learner algorithm, and team 2 developed an algorithm based on the recursive least squares methodology.
After two days of intense work followed by a dinner graciously offered by the Société de Réanimation de Langue Française (SRLF) and a party till the end of the night in the Quartier Latin of Paris, ACTERRÉA team 1 won the 3rd prize! Team 1 represented by Ményssa Chérifa presented their research at the international congress of intensive care organised by SRLF on the 24th of January.
Well done to them all!
-- refs --
MIMIC-III, a freely accessible critical care database. Johnson AEW, Pollard TJ, Shen L, Lehman L, Feng M, Ghassemi M, Moody B, Szolovits P, Celi LA, and Mark RG. Scientific Data (2016). link
Pollard, T. J. & Johnson, A. E. W. The MIMIC-III Clinical Database (2016). link