Predicting hospitals resources & their upcoming emergencies
Interview with PhD Hugo Flayac, Head of research at Calyps
Following the last episode of Airccelerate podcast discussing the mediCAL.ai project with Tony Germini, CEO of Calyps, our association had the chance to interview Hugo Flayac, Head of research, on the genesis of the project.
Portrait of Hugo Flayac
Hugo Flayac began his career as a doctor in quantum physics: after several years of research at EPFL, he realized that data analysis and artificial intelligence (AI) were becoming increasingly important in the profession. He became interested in AI and trained with experts at deeplearning.ai. In 2018 the engineering school of HEIG-VD opened a position for an Innosuisse project combining two of his passions: AI and health data. He jumped at the opportunity and took the lead of the research project. Planned to last 18 months and launched in partnership with Swiss-based company Calyps, the aim of the project was to optimise ambulatory flows in a hospital using machine learning (ML) techniques. The prediction of key indicators such as “no-show” (patients or practitioners failing to show up for the planned operation), the duration of medical operations as well as the probability of patient death made of this first stage a success. Since 2019, Hugo is fully employed at Calyps and leads research for the company.
Emergencies overcrowding in France
For sure, emergencies are difficult to predict: their numbers are unstable, require frequent and unforeseen adaptations of either health care staff, or medical facilities. Knowing that a third of visitors registering at the emergencies will be hospitalized, anticipating the people flow — patients, 65+ old patients, health care staff — would be a must.
From the initial discussions between Rodolphe Bourret, general manager of the Centre Hospitalier de Valenciennes (CHV), and Tony Germini at the HIT 2019 conference, to the lengthy discussions between Calyps and CHV employees, all tend towards the same conclusion: overcrowded emergency departments as experienced in most of French hospitals are slowing down a significant part of the medical activities led in other vital departments.
From problem to solution
Initiated in Switzerland, the first stage of the mediCAL.ai prototyping involved the Ensemble Hospitalier de la Côte (EHC), providing a large volume of data related to the planning of hospital operations. After conclusive results on outpatient operations, attention was turned to the hospital’s emergency departments: a collaboration was quickly agreed in France at CHV between Dr. Maisonneuve’s team, the hospital’s IT department and Calyps’ data experts. Patient data is stored in CHV data centre, where it is anonymized and regularly transferred to Calyps.
CHV is a large hospital with 1,800+ beds and around 200 emergency admissions per day. Thanks to this large volume of statistical inputs we were able to calculate predictions for Emergencies arrivals with a low margin of error.
These data enabled Calyps to build ML models predicting the number of patients who will be admitted to the hospital’s emergency department up to 7 days in the future. From a success rate of 85% in the first versions, these models have achieved predictions with a reliability rate of over 90% after training and adjustments. With data updated every hour, the model can learn continuously (see reinforcement learning explained in the interview of Karim Bensaci, Head of products & solutions) and improve over time.
The mechanics of the system
Contrary to what you may think, the system is not a black box, rather a sequence of several modules, each one doing its part. One module will estimate how long the patient will wait in the emergency desk, another will predict whether the patient would be transferred in another department or leave the hospital, and so on. The data used by the system are of three types:
- Patient-specific data such as age, weight, height, diagnosis on arrival, vital signs and more.
- Volumetric data, which are time series indicating, for example, the number of patients per hour at the Emergencies.
- Environmental data such as weather forecasts (or major events in the region, less relevant for 2020).
Regarding environmental data, two anecdotes are worth highlighting:
- Social media did not have a significant impact on the predictions given by mediCAL.ai, as they were not that used in Valenciennes (ca. 45'000 inhabitants) and its surroundings (urban area less than 400'000 inhabitants).
- The matches played by the Valenciennes Football Club until 2019 could gather more than 20'000 spectators at the Stade du Hainaut. Each match brought its share of injured people, five to ten of them arriving at CHV after each soccer match, as doctors remember!
Regarding patient-specific data, the technologies used predictions are standard neural networks, also called feedforward neural networks. Whereas for volumetric data, another type of network had to be used in order to take into account the seasonality of data (i.e. there are more patients on Mondays than during the rest of the week, a peak in admissions is usually felt between 2 and 3 p.m., patterns of weeks repeat, patterns of seasons repeat, and so on). To solve this problem, Calyps uses convolutional neural networks, or “CNN”: this type of network is inspired by the visual cortex of animals and has proved its worth in image analysis.
COVID-19, kryptonite of AI models
Artificial intelligence algorithms learn from past data in order to “predict” the future. An unexpected event such as this pandemic has put a strain on the mediCAL.ai system, which never saw this before. “COVID-19 is the kryptonite of AI models”, says Hugo. However, his system can now learn from current data to predict an upcoming crisis, up to a certain point.
What happens next
The mediCAL.ai system is currently proving its worth at CHV. The problem seems less critical in Switzerland, even if this innovation is already attracting interest of hospitals and clinics. For the latter, the planning of outpatient operations should not be postponed and the medical coding necessary for invoicing an operation could benefit from automatic precoding.
This article was co-authored by Marco Brienza and me.
