Analyse multimodale de la parole pour la détection précoce des poussées de la maladie de Crohn grâce à des méthodologies d'apprentissage profond.

Keywords

Artificial Intelligence, Deep Learning, Flare detection, Facial expressions

Subject details

** Context ** Inflammatory bowel diseases (IBD), particularly Crohn's disease (CD), pose significant challenges to healthcare systems and patients due to their chronic and unpredictable nature (Strober et al. 2007). CD affects a substantial portion of the population in France and Europe, with considerable impacts on patients' quality of life and healthcare resources. Within the framework of the I-DEAL project, the main goal is to address the unmet needs of CD patients by enabling early intervention and promoting a return to a normal life through innovative remote monitoring solutions. One of the objectives of the I-DEAL project is to develop a home-based system for early detection of CD flares, that enable timely intervention and improve the management of CD patients, ultimately enhancing their quality of life. ** Objective of the Thesis ** The objective of the thesis is to characterize pain flares through facial expressions and modifications in speech patterns using NLP, machine learning, and computer vision techniques. This will be achieved by analyzing video recordings of multiple patients, with each recording annotated by a physician to indicate the patient's condition at the time of recording. The goal is to extract features from these recordings that capture the nuances of pain flares, enabling the development of classification models capable of identifying and classifying pain flare states accurately. ** Missions ** The missions of the PhD encompass several key scientific objectives. Firstly, active participation in the collection of audiovisual data from patients, with a specific focus on eliciting the pronunciation of predefined sentences. Secondly, employing advanced techniques for feature extraction, particularly linguistic features derived through sentiment analysis, while simultaneously annotating the dataset with pertinent medical information. Thirdly, the utilization deep learning methodologies, such as ResNet CNN, BLSTMs) (Tsai et al. 2017) or Time delay neural network (TDNN), to facilitate the training of classification models on the processed data (Fontaine et al. 2022, De Sario et al. 2023, Othman et al. 2021, Littlewort et al. 2007). Finally, an integral aspect involves evaluating the efficacy of the developed methodology by rigorously analyzing model performance and validating its capacity to accurately discern levels of Crohn's Disease severity based on speech and visual modalities.