Machine Learning-Based Prediction of Inpatient Length of Stay at CNRFR – Rehazenter Rehabilitation Center

Luis Dierick, Benoit Frenay

Track: Type D (Master/Bachelor Thesis Abstracts)

Keywords: Length of Stay, Rehabilitation, Machine Learning, eXplainable Artificial Intelligence

Abstract:

Length of Stay (LoS) is a crucial metric in healthcare, reflecting patient progress and informing
resource planning. While LoS prediction has been widely studied in acute care settings, its
application in rehabilitation remains limited. This thesis investigates the potential of Machine
Learning (ML) models to predict LoS at admission in a rehabilitation center using real-world data
from the Centre National de Rééducation Fonctionnelle et Réadaptation (CNRFR – Rehazenter)
in Luxembourg.
This work explores multiple preprocessing strategies, regression algorithms, and techniques to
address the skewed nature of LoS data. A particular emphasis is placed on model explainability,
supported by SHapley Additive exPlanations (SHAP) to ensure clinical interpretability. The
study considers two distinct patient groups: neurological and traumatological.
Results show that ensemble tree-based methods, particularly Random Forest (RF) and Categorical
Boosting (CatBoost), outperform Multiple Linear Regression (MLR), achieving an $R^2$ of up
to 0.51 in the traumatological group and 0.25 in the neurological one. Key predictive features
for LoS include Functional Independence Measure (FIM) and number of chronic diseases for
both groups, with additional contributions from socioeconomic and comorbidity variables in
neurological patients.

Serve As Reviewer: ~Benoit_Frenay1

Submission Number: 56