Title: Global method for predicting the hospital length of stay using incremental and evolutionary data.
Supervisor: Professor Antoine Duclos.
Abstract: Predicting patient length of stay is an important issue for the organization of care activities in hospitals, especially for beds management the and preparation for patients discharge. Facilitating the organization of hospital activities influences access, quality and efficiency of care. In this thesis, we sought to predict length of stay for all patients in the hospital, at all stages that make up their care pathways, using standardized Medical, Surgical, Obstetric medico-administrative data collected for reimbursement of care.
We began by conducting a systematic review of the literature on methods for predicting lengths of stay, in order to better understand data preparation, the different prediction approaches, and how to report the results. We then worked on a data preprocessing method and investigated the ability of embeddings to represent medical concepts in the context of length of stay predictions via a neural network. The ability of the neural network to correctly predict length of stay was rigorously evaluated and compared with a random forest and a logistic regression. This work shows that hospital length of stay can be predicted by a neural network using standardized medical-administrative data available for all patients.
- Boyer, Laurent, PU-PH, Université Aix-Marseille, Rapporteur
- Bringay, Sandra, PU, Université Paul Valéry Montpellier III, Rapporteur
- Soualmia, Lina, MCF, Université de Rouen, Rapporteur
- Fromont, Elisa, PU, Université de Rennes, Reviewer
- Schott-Pethelaz, Anne-Marie, PU-PH, Université Claude Bernard Lyon 1, Reviewer
- Duclos, Antoine, PU-PH, Université Claude Bernard Lyon 1, Supervisor
- Wang, Tao, MCF, Université Jean Monnet Saint-Etienne, Co-supervisor
- Fondrevelle, Julien, MCF, Institut National des Sciences Appliquées Lyon, Co-supervisor