Hospitals are brought to optimize the use of their resources to meet the growing health needs of the population and many budget constraints. Estimated length of stay (LOS) should be established at each admission to inpatient unit in order to plan care activities appropriately. Inaccurate estimation can make the organization of hospital activities inefficient, lead to overworking of health professionals and long waiting time for patients. In this context, the aim of the thesis topic is to develop a new prediction method that will assist the bed manager in estimating the LOS at the moment of inpatient admission and during the hospital stay.
The prediction of LOS is a complex problem that depends on many factors related to patient’s clinical and social context, his care treatment and care unit in which the patient is admitted. The answer of an empirical prediction is not always valid for lack of information or experience, especially in the case of non-frequent inpatient stays. The results of our first study, conducted through a HeSPeR-HCL-INSA teamwork and carried out with statistical methods, showed that the prediction performance was excellent for LOS less than 2 days, but hardly exploitable for long LOS. However, the predictive ability of these methods was superior to that of humans according to a study conducted with 25 physicians on a sample of 187 stays.
The investigation must be pursued to improve the accuracy of predictions, by seeking new methods from Big Data and Artificial Intelligence. It seemed appropriate to use machine learning algorithms with the data routinely collected in the hospital's information system. These algorithms allow the computer to extract implicit knowledge from the masses of data.
The challenge is to design a global method which is (1) appropriate to select predictive information qualifying patient's profile, (2) adapted to incremental and scalable data collected during patient’s care pathway, and (3) able to garner experiences and knowledge from medical data containing elements of uncertainty and hesitation.
This topic takes part in the establishment policy of the Hospices Civils de Lyon in terms of performance and relevance of care. The development of the prediction method will use the information available for patients hospitalized at HCL between 2011 and 2018. The prediction method will easily be generalized to other healthcare institutions, it will contribute to improve the quality of care provided in Auvergne-Rhône-Alpes Region and the efficiency of hospitals without requiring additional human resources to predict the LOS.