Applying machine learning and AI to predict patient risk

To reduce adverse health events, a leading hospital in France teamed up with Elsevier’s data scientists

By Alice Mogenet - May 9, 2022  4 mins
CHUGA collage
To reduce adverse events at Centre Hospitalier Universitaire Grenoble-Alpes (CHUGA) in France, Head of Public Health Prof Jean-Luc Bosson, MD (upper right), worked with Elsevier data science experts, led by Sigurd Prieur, PhD, VP of Life Science Analytics.

Healthcare, by definition, is supposed to make you better. But sometimes an infection is contracted at the hospital or a complication occurs after surgery. Such health-related adverse events occur in 8% to 12% of all hospitalizations, according to the World Health Organization. There are 750,000 health-related adverse events in the European Union each year, which amounts to more than 3.2 million days of hospitalization that could have been prevented, WHO reports.

A 2017 report from the Organisation for Economic Co-operation and Development (OECD) shows that more than 10% of hospital spending is related to the treatment of health-related adverse events that occur during hospitalizations.

The desire to improve this situation sparked a collaboration between Centre Hospitalier Universitaire Grenoble-Alpes (CHUGA) in France and Elsevier.

In 2019, Prof Jean-Luc Bosson, MD, head of the hospital’s Public Health Department, teamed up with Elsevier to apply machine learning to their historical patient data with the aim of creating models that identify patients at higher risk for healthcare-related adverse events. To do this, they needed to build a single multisource dataset, or “data warehouse,” that combined all the hospital’s internal data sources. This complex task involves sourcing data from different places, such as laboratories and radiology departments, and incorporating various types of sources, such as diagnoses, notes and orders from nurses and physicians. The task also requires resolving data mismatches and coding inconsistencies.

Over the course of the pandemic, the project team from CHUGA worked with Elsevier colleagues remotely to set up the pre-conditions for big data analysis using modern machine learning methods. Prof Bosson remarked:

This type of project benefits the patients first, but also the organization. Before, I viewed Elsevier essentially as a publisher of scientific journals. With this project, I discovered and understood Elsevier’s openness to a world we share — medical informatics and health analytics.

Prof Bosson’s ambition is now almost in place: simultaneous modeling of hundreds of variables to uncover relationships, look for patterns and define populations at risk. The models identify the top 5% of patients with a 4.7x increased risk for life threatening event lik thromboembolism or a 40% risk of a prolonged hospital stay. This will allow the hospital to flag patients that fit the risk profile and provide more directed care.

Without Elsevier’s data science teams, we would never have had the expertise and availability of sufficient staff to complete this project. Or we would have done it in five to six years, and the project would have been obsolete before it was finished. In this field, you have to have quick results because things move very fast.

A cross-disciplinary project team

Ursula von Schenck, MD, PhDFrom his base in Berlin, Sigurd Prieur, PhD, VP of Life Science Analytics at Elsevier, led the international project. Elsevier’s Senior Medical Officer, Ursula von Schenck, MD, PhD, worked closely with Prof Bosson and Prof Daniel Pagonis on the conceptual design and the interpretation of the analytic results.

Elsevier’s lead data scientists developed a sophisticated toolbox to pre-process, model and validate big data such as electronic health records utilizing state-of-the-art machine learning and validation algorithms:

Combining our own expertise with that of our customers enables us to come up with customized solutions that meet their needs and improve patient outcomes. As Dr Prieur explained:

Elsevier is increasingly positioned at point-of-care decision science, where we combine data and content to help streamline the care process in hospitals. This project demonstrates our ability to understand a data stream, gain insights from it and make something better, all the while respecting data privacy and GDPR compliance.

A predictive model

DEMETER spider
Machine learning models enable clinicians to predict a patient’s individual risk for healthcare-related adverse events at admission to target resources efficiently and improve the patient’s outcome. This predictive model is based on lab values, procedures, diagnoses, social determinants of health, and entry and exit mode. (Source: Result from DEMETER, a retrospective and observatory study between CHUGA and Elsevier.)

This story is based on a case study that appeared in the 2021 RELX Annual Report.

Contributors


Alice Mogenet
Written by

Alice Mogenet

Written by

Alice Mogenet

Alice Mogenet is International Senior Product Manager for Clinical Decision Support Solutions at Elsevier. She is based in Sèvres, Île-de-France, France.

Related stories


Together we created the most advanced 3D female anatomy model
3 ways AI is shaping the future of clinical research

Comments


comments powered by Disqus