How we revolutionize the operational management of hospitals with (artificial) intelligence
Are you already involved in an artificial intelligence (AI) business ? Although I have been analyzing data for more than 20 years, my company CALYPS entered the new era of data science and AI quite recently.
Our team has indeed developed an AI able to relieve congestion in a hospital emergency room (ER). And it works: simply named Calyps AI, our software currently in production at France-based Valenciennes Hospital (1800+ beds) anticipates the number of ER arrivals up to 5 days before, with a confidence level up to 90%. It also integrates a bed dispatching tool for each key department, initiating a revolution in terms of case predictions and resource allocation. Even national TV channels (cf. news flashes on France 2, France 3 or France 24) praised the relevance of Calyps AI’s predictions… what an honor !
So how did we build our dream team — Karim Bensaci, products & solutions manager ; PhD Hugo Flayac, data scientist ; Abdourahmane Faye, cloud & AI platforms expert — and such a disruptive healthtech solution in less than 2 years ? Let’s go back to the future.
(Y-5) Before the acceleration phase
In 2015, CALYPS began its strategic shift by conducting its first study in Switzerland amongst various representatives and employees of public hospitals, or private clinics. Our survey was related to the new “medical-economic” data in use at that time: for this purpose, we interviewed 30 medical executives, CEO’s, CTO’s, chairmen and board members.
In answer to the need for more KPIs we developed an analytical platform called mediCAL dashboard (demo) whose objective is to centralize data provided by the information systems of a healthcare institution (patient file, RIS, laboratory, etc.) screening a 360° view of the patient journey.
In 2017, after implementing the mediCAL dashboard to optimize the organization and the planning of ambulatory operations in a Swiss hospital, we realized that we could do more than just represent the past. We had data available in the form of time series, well suited to the use of Machine Learning (ML).
(Y-2) Acceleration & findings
With this in mind, we started an InnoSuisse research in 2018 in collaboration with the HEIG-VD (University of Applied Sciences) and one of their most brilliant data scientist, PhD Hugo Flayac. Our aim was to develop a solution for the management and planning of ambulatory operation flows under uncertainty.
We collaborated with a hospital in the Lake Geneva region and, after a few months, switched from ambulatory to in-patient flows as the need for planning under uncertainty was even more obvious. In this project, we were able to transcribe the reality of the field by analyzing its data. I still keep in mind the quote of a management board member:
‘Incredible, with little input from our part you were able to represent the reality in such a short time’
We were therefore able to provide elements of anticipation for ambulatory and inpatient flows, notably through the prediction of length of stay. In a discussion with a consortium of Parisian hospitals, we understood that uncertainty was at its height when ER arrivals had to be integrated into a planned flow. As emergencies generate around 30% of hospitalizations over a year, they will somehow disrupt the organization and care planning.
(2019) First implementation of Calyps AI
Our implementation at Valenciennes General Hospital was done gradually, starting with the integration of hospital data. First things first: data preparation. At each iteration, the objective was to evaluate the variables that impact predictions and to take them into account in our predictive model. Only then did we integrate data from outside the hospital such as weather, sports venues — i.e. soccer matches at mythic Hainaut stadium can also lead to their load of injuries — cultural events, holidays or vacation schedules.
We proceeded hand in hand with the medical staff who had to interpret and act on these predictions. Internal collaboration (1) evangelization (2) and training (3) are three key elements of adherence to the system, building trust in the predictions and allaying any residual fear.
During summer 2019, we started a real time experiment at Valenciennes Hospital and, six weeks later, provided daily predictions of emergency arrivals with an accuracy of >90% over the period.
How Calyps AI works
In brief, whether it is an emergency arrival flow or the patient trajectory, data needed for good predictions is more or less the same. Thanks to this, the chain of actions related to the flow can be anticipated and has a positive impact on both, the caregivers and the patients.
Our algorithms are based on artificial neural networks (Deep Learning) fed by a variety of past and present data provided by the hospital. We build signals considered as key and train the system to find linkages between these signals, with a sense of time. For example, admissions clearly have an impact on hospitalizations, the department load (the related indicator is called Tension) and hospital discharges to follow up. On the other hand, admissions are influenced by external factors: weather, holidays, events, seasonal illnesses which must be integrated to reach stable performance throughout the year, particularly on peak activities.
Our policy is to inform without deciding at the doctor’s place: admission flow forecasts allow the Chief of Emergency Services to size his teams optimally 5 days in advance. Over a day, our Tension forecasts (number of patients present at a given time) enable each team to better anticipate the department’s situation hour by hour. We then provide accommodation solutions (bed) downstream of the emergencies for patients requiring hospitalization.
Calyps AI alive!
Look at the 3 min. demo of Calyps AI here.
Currently, Calyps AI is used daily by the Chief of Emergency Services at Valenciennes Hospital. Doctors and caregivers also get their own level of information (chief gets predictions at 5 days, doctors at 2 days, caregivers at 2 hours). Finally, the benefits are many: staff are relieved in case of major tension and can focus on their primary mission. Hopefully, patients experience this indirectly.
From flow anticipation in ER to downstream flow management and post-hospitalization
Being able to predict incoming flows, their typology and duration, the link with overall hospital planning and patient care seems obvious. Since the beginning of 2020, we have been working on the next generation of Calyps AI and are planning the deployment of a hospital coordination platform with the objective of real-time and smart bed management including recommendations for stays, site allocation and anticipation of discharges.
Regarding Covid-19, we built specific signals to detect patients with flu-like symptoms and therefore potentially Coronavirus. Our algorithm still learns from the current situation and “draws inspiration” from the impacts of the flu to anticipate cases of Coronavirus. In case of new outbreak (or the resumption of a peak) our system will be ready to better anticipate. Real-time territorial data would be an additional asset to refine our predictions in connection with Covid-19 and we are actively working on this, notably in France.
A story to be continued for sure.
