Since it is would be extremely unusual that one individual would have both the clinical training and the computer science background for BDA and ML, teams with professionals across different fields are required, composed of data scientists including statisticians and epidemiologists, clinicians and other biomedical researchers. The coordination of a group with different schedules, the establishment of clear objectives and an effective leadership to manage members with different cultures and who “speak different languages” are all individual challenges in themselves,11 although in critical care medicine teamwork is routinely applied in the day-to-day workflow.

This article describes an initiative event organized by the Massachusetts Institute of Technology (MIT) Critical Data group, the Life Supporting Technologies (LifeSTech) research group at Universidad Politécnica de Madrid (UPM) and the Innovation Unit and Critical Care Department of Hospital Clínico of San Carlos (HCSC) in Madrid.

The use of BDA and ML techniques creates a new avenue for clinical research that includes non-traditional, non-academic investigators, especially when large anonymized databases of critical patients are shared under pre-specified data user agreement. The pioneers in this field are Beth Israel Deaconess Medical Center and MIT, whose partnership created the MIMIC database (Medical Information Mart in Intensive Care, http://mimic.physionet.org) for the research community. The MIMIC-III version contains data from more than 50,000 admissions to Beth Israel ICUs from the years 2001 to 2012 and is available online under a data user agreement. It includes vital signs and other monitoring data, laboratory results, data of prescription and administration of drugs, as well as progress notes, diagnostic codes, physiologic waveforms, both in the ICU and for some variables, from throughout the entire stay in the hospital.

Similar high-resolution ICU databases exist, like the Critical Care – Health Informatics Collaborative (CC-HIC) Database,12 but MIMIC is unique because it is completely open-access. In addition, the MIT group published a textbook that is available at no cost and teaches a course to guide investigators on the analysis of the database.3,13–15 Since 2016, they have also organized several datathons16 in various regions of the world (MIT, London, Paris, Melbourne, Beijing, Sao Paulo and Singapore), to jumpstart disseminate the field of machine learning as applied to health data.

The creators of MIMIC have proposed a system of learning and continuous improvement in Intensive Care they call “closed circuit”.8 This strategy consists in using the patient database to answer questions of clinical interest and then taking it back to the bedside.

Although adoption of electronic medical records and databases of critically ill patients is still in the growth phase in our country, and not all Units have implemented an integrated digital information system, there are already some initiatives under way.

The Spanish Society of Critical Care (Sociedad Española de Medicina Intensiva, Crítica y Unidades Coronarias, SEMICYUC) has published reference standards for the implementation of Clinical Information Systems (CIS),17 and is in the process of building the national electronic record of the minimum basic data set (CMBD) registry to facilitate quality measurement, benchmarking and clinical research.

In some services and hospitals, the process of creating anonymous data bases for collaborative research has already begun. The Critical Care Department at Hospital Clínico San Carlos in Madrid, in collaboration with Philips, has already anonymized its database including its first 12,000 admissions after implementation of the IntelliSpace Critical Care and Anesthesia (ICCA™, Philips) clinical information system. Other ICUs in Spain, such as the Unit at Hospital Joan XXIII in Tarragona, are working on similar initiatives, using the data obtained automatically from their CIS.18,19

The creation and research use of these patient-level databases have some important legal implications, which need to be addressed to comply with regulatory policies, particularly those under the Spanish Data Protection Law. Projects that employ BDA and ML of these patient databases also require prior evaluation and approval by local Ethics Committees.

In order to promote the use of BDA and ML in healthcare in our country, the seventh Critical Care Datathon, organized jointly by the MIT Critical Data Group, the HCSC in Madrid and the LifeSTech from UPM, took place on December 1, 2 and 3, 2017 at Impact-Hub co-working space in Madrid. The event brought together clinicians, computer scientists and other professionals interested in health data science. HCSC and LifeSTech-UPM group collaborated in the event diffusion informing participants about the content and meaning of the Datathon.

Prior to the event, a group of clinicians from the Critical Care Department of Hospital Clínico San Carlos prepared a list of eight research questions of clinical interest, by way of brain storming, discussion and reaching internal agreement, to be explored during the event: (1) the choice of empirical antibiotics in the ICU on hospital mortality, (2) risk factors for prolonged ICU stay and mortality after cardiac surgery, (3) prediction of poor prognosis after admission for resuscitated out-of-hospital cardiac arrest, (4) prediction of mortality in patients with non-elective ICU admission, (5) prognostic implications of hyperglycemia in critically ill patients with and without diabetes, (6) the epidemiology of massive transfusion of packed blood cells, (7) understanding the association of positive fluid balance and mortality, and (8) seasonality of subarachnoid hemorrhage.

After brief presentations of the background and the rationale behind the research questions with consideration of the non-healthcare professionals among the Datathon attendees, eight groups were formed, one for each topic, with the stipulation that all the necessary expertise (clinical, data extraction and processing, machine learning) is represented in each group.

The MIT Critical Group with the support of LifeSTech from UPM set up for the event virtual servers with the MIMIC-III database with access to platforms and software for data analysis (Python, R, Jupyter notebook). LifeSTech evaluated using natural language processing techniques to extract information from clinician's reports written in free text, with the purpose of codifying them through SNOMED-CT terminology.20 Mentors from the MIT team and LifeSTech were on hand throughout the weekend to help troubleshoot issues, both technical and those related to the MIMIC database or to data modeling. Finally, each group presented their methods and results to the rest of attendees and a panel of judges from MIT, UPM and HCSC.

As an illustration of the relevant conclusion rendered by the collaborative work of the groups during the datathon, we briefly describe the experience of the group who worked on the project looking at the prediction of mortality among patients with non-elective ICU admission.

First, the clinician in the team presented the existing critical care mortality prediction models (APACHE II, MPM, SAPS 2, SAPS 3), and interest in developing an updated score with better calibration from a larger database. A literature search was carried out to select the scores with the best performance for comparison. The OASIS21 and APACHE II22 scores were chosen as the current “gold standard”. OASIS, which was constructed using ML techniques, has the best performance among critically ill patients in the MIMIC III database (area under the receiver operating curve or AUROC, 0.82) (personal communication from MIMIC team), while the APACHE II score, the most frequently used for benchmarking and clinical research, has an AUROC of 0.75.23

Likewise, the clinician identified variables in MIMIC-III potentially related to prognosis. The data scientists were responsible for generating the table of more than 40,000 patients with all the selected variables. The entire team went over phases of the model building. In the first phase (training), Support Vector Machines (SVM) was trained using 25,000 patients and produced a model of 30 variables with an AUROC of 0.81. This performance was validated in a different subset of 15,000 patients, with a similar AUROC, confirming the robustness of the model. Presently, the group is working in refining the model and evaluate its performance in the anonymized ICCA™ (Philips) database with 12,000 patients of our Critical Care Department.

The same collaborative work between clinicians and data scientists performed in previous similar events and that we witnessed in the Madrid Datathon, has led to numerous publications,24–28 and we have the same expectation from the Madrid event.

The meeting, which was organized under the auspices of SEMICYUC, also served to connect Spanish ICUs working in the field of BDA and fortified partnerships among them to move this initiative forward.

We are conceivably at the dawn of a new era in intensive care, in which clinical decision making will increasingly be assisted by computers that perform data integration and analysis. The clinician will therefore have to navigate a specialty of critical care, that harnesses the power of data to individualize care in order to improve population health, where collaborative work with other non-healthcare specialists in the area of data science is essential to leverage all the information that is routinely collected in the process of care, but without compromising patient rights to privacy.

We encourage all those interested in participating in this project to contact the authors, in order to establish a national collaborative network in the field of BDA and ML as applied to healthcare data collected in the ICU, under the auspices of SEMICYUC.

MIMIC III experts: Celi, Armengol, Deliberato, Paik, Pollard, Raffa, Torres.

Technical logistics (servers and database access): Celi, Armengol, Deliberato, Paik, Pollard, Raffa, Torres, Mayol, Chafer, Rey, Gonzalez, Fico, Lombroni, Hernandez, Lopez, Merino, Cabrera, Arredondo.

Preparation and presentation of research questions: Blesa, Martín, Nieto, Martínez, Álvarez, Sánchez, del Pino, Gil, Núñez.

Clinical support for the research questions: Sanchez, Núñez, Álvarez, Martínez, del Pino, Gil, Bodí, Rodríguez, Gómez, Prada.

Concept and Design of the Datathon: Celi, Armengol, Deliberato, Paik, Pollard, Raffa, Torres, Mayol, Chafer, Rey, Gonzalez, Fico, Lombroni, Hernandez, Lopez, Merino, Cabrera, Arredondo.

Drafting of the manuscript: Núñez, Sánchez, Celi, Bodí, Fico, López.

Venue of the event: Gonzalez, Mayol, Chafer, Rey.

Diffusion: Chafer, Rey, González, Sánchez.

The authors declare no conflict of interests.