用于神经退行性疾病声学生物标志物的识别与可解释性研究

本项目旨在利用机器学习、声学与语言分析以及可解释人工智能（XAI）方法，识别并解释与神经退行性疾病相关的语音生物标志物。它将重点开发可解释的模型，通过利用与各种疾病相关的语音数据集，协助早期检测与监测工作。

Job description: Project description: Medicine is becoming increasingly personalized, driven by the use of longitudinal data collected throughout the course of care. This wealth of information has fostered the adoption of machine learning models as decision-support tools capable of predicting a patient's status, estimating the risk of progression toward a clinical event, and identifying factors associated with diverse individual trajectories. However, the translation of these models into clinical practice remains constrained by inter- and intra-individual variability, as well as the heterogeneity, incompleteness, and temporal irregularity of the data—factors that complicate the development of robust predictive models. These challenges are particularly pronounced in the case of neurodegenerative diseases, where early detection and the monitoring of clinical progression remain major hurdles, largely due to heterogeneous individual trajectories and follow-up data that are often incomplete or temporally irregular. In this context, speech serves as a non-invasive and easily accessible data source capable of providing digital biomarkers that reflect motor, cognitive, or executive impairments. Automated analysis allows for the combined use of acoustic and linguistic features to characterize various pathologies and track their progression. Yet, identifying robust vocal biomarkers remains an open scientific challenge. A feature strongly associated with a prediction is not necessarily a generalizable biomarker; it may be influenced by factors such as age, gender, the speaker, the speech task, or recording conditions. This issue becomes even more complex in a longitudinal setting: a feature that effectively distinguishes between two groups at a specific point in time may not necessarily provide information regarding disease progression. Consequently, it is essential to simultaneously examine the patient's current status, variations between sessions, and the individual trajectories of the biomarkers. Machine learning models enable the utilization of this complex data, yet the highest-performing models often remain difficult to interpret. Explainable Artificial Intelligence (XAI) methods help identify the factors contributing to predictions, but their outputs are typically expressed in technical, quantitative terms that are not easily actionable for clinicians. The emergence of Large Language Models (LLMs) offers the potential to transform quantitative XAI outputs into natural language explanations that are contextualized, readable, and interactive. However, this approach raises questions regarding fidelity and traceability: the reformulation must remain consistent with the model's underlying explanation, without introducing unsupported information, spurious causal relationships, or erroneous interpretations. This thesis aims to develop an integrated approach to identify, characterize, and explain vocal biomarkers associated with neurodegenerative diseases by combining acoustic and linguistic analysis, machine learning, longitudinal modeling, and XAI methods. The LLM operates downstream of the predictive model to translate and explore the explanations generated by XAI methods. The research will draw upon several speech datasets linked to various neurodegenerative diseases, including Parkinson’s disease, Alzheimer’s disease, and other forms of dementia affecting speech. Particular attention will be paid to identifying markers that are either common to or specific to different pathologies, as well as assessing their robustness against sources of variability and their evolution over time. Envoyer des commentaires

Profile and skills required: Essential skills, knowledge and experience: - Deep learning methods - Data processing and classification methods - Basics in signal processing - Languages Python, Matlab, C, Unix,…. - Knowledge of scientific library software specific to Machine Learning, Signal and Data Processing (Pytorch, KERAS, etc.) - Knowledge in the domain of LLM Abilities and skills: - Fluency in English (written and spoken) and French (very apreciated) - Teamwork ability - Analytical skills and rigor - Ability to tackle cross-disciplinary scientific subjects (AI, signal processing and data exploitation) - Ability to work in a national and international scientific environment Niveau de français requis: Compétent/Courant: Vous pouvez utiliser la langue avec aisance et facilité en argumentant sur des sujets complexes. Niveau d'anglais requis: Intermédiaire supérieur: Vous pouvez utiliser la langue de manière efficace et vous exprimer précisément.