Machine Learning for DSS

The project was carried out as part of Ms Racda Naïla MEKHALDI's PhD thesis. It is the fruit of research work  carried out at LAMIH from October 2018 to January 2022. This research is based on field knowledge provided by ALICANTE, a company specializing in the development of tools dedicated to the hospital environment.

In recent years, healthcare establishments have been constantly seeking to optimize the operation of their services while ensuring the quality of these services. Hospital Length Of Stay (LOS) is an indicator for evaluating the performance of care facilities and the efficiency of hospital services. As a result, the estimation of a patient's DDS not only on admission to a service, but also throughout the course of his or her care has been the subject of several studies.

The prediction of DDS contributes to optimizing the use of hospital resources, improving the organization of care through better planning of activities.

A literature review was carried out to identify the various DDS models existing in a hospital environment. We then deduced a generic model characterizing DDS in several medical units by adding new information defined based on daily hospital needs. Our approach to DDS prediction is based mainly on machine learning and data mining techniques. Two prediction models have been proposed.

The first model is static. It aims to determine the DDS as soon as a patient is admitted to hospital. For this, available data on previous admissions and "machine learning" tools were exploited. The work focused, in part, on data formatting, "cleaning", completion, normalization etc.

The second model is dynamic in nature. It aims to enrich the first model with data that are acquired during the patient's length of stay. The dynamicity of the model is sequential. The DDS history is seen as a series of dated data or events that need to be linked in a coherent way.

The contribution of this project does not concern the definition of new machine learning algorithms  but above all their exploitation in the hospital field. Indeed, the complexity of medical data is a major difficulty that needs to be addressed. This complexity stems from the great variability in the origin of the data and their low level of formalism (free handwritten records). Data from the Programme de Médicalisation des Systèmes d'Informations (PMSI) were also used to put our contributions into practice.