Trustworthy AI for hospital length of stay prediction

Internship period: Feb to July / 2026 

Funding: DISP lab 

Required profil: master’s degree in computer science (or equivalent, such as an engineering school), Analytical mindset, Curiosity.

Skills: Python, Tensorflow, Deep Learning 

Length of stay (LOS) prediction accuracy is critical for hospital management and bed capacity planning, which influences healthcare delivery, quality and efficiency. The goal is to predict LOS using advanced trustworthy Machine Learning on administrative data from acute and emergency care and compare this method to conventional methods. A model leveraging embeddings and a feedforward neural network (FFNN) was developed to provide accurate LOS predictions at each stage of hospitalization. Its performance was compared against a random forest and logistic regression using metrics such as accuracy, Cohen's kappa, and a Bland-Altman plot. The FFNN achieved a prediction accuracy exceeding 94% for stay durations ranging from 0 to 14 days. 
However, its applicability and explainability become an issue when considering real-world clinical settings. While FFNN demonstrates high predictive accuracy, its black-box nature poses challenges for healthcare professionals who require interpretable insights to make informed decisions. This lack of transparency may hinder its acceptance and integration into clinical workflows.
The main objective is to draw inspiration from the achieved results and focus on improving the model's exploitability while enhancing its trustworthiness for clinical applications. This involves developing strategies to make predictions more interpretable and actionable for healthcare providers. Techniques such as feature importance analysis, surrogate models, or integrating domain knowledge into the model's architecture can bridge the gap between high performance and usability. Therefore, the candidate models should be extended and challenged with performance metrics about bias, fairness, explainability, and Trustworthy. These metrics should be justified in their definition, selection, and integration at the different steps of ML model’s design.
Furthermore, the goal is to refine the development approach so that it not only predicts outcomes accurately but also provides insights that align with clinical reasoning and ethical considerations. 
 

Proposed Work Plan: 
•    Literature review on LOS prediction, HIS, resource management, AI performance metrics (1 month)
•    Development of approaches (2 months)
•    Experiments and validation (2 months)
•    Writing a scientific article to highlight the achieved results (in parallel, final 3 months)
 

Submit your application by email to sara.bouguelia(AT)univ-lyon2.fr, including a file composed of a CV, recent transcripts, a cover letter, and a recommendation letter.