Manuel Mounir Demetry Thomasen

Position: Development of vocal biomarkers for predicting diabetes-related complications using deep learning
Categories: Cross-Academy Fellows 2024, Fellows
Location: Aarhus University

Abstract:

Background Diabetes is a complex chronic condition with severe potential complications, which poses a huge burden on people with the condition, their families, and the healthcare sector. Risk assessment tools facilitating early detection of complications are crucial for prevention and progression management. Progression of diabetes and corresponding physiological changes affect several organs involved in the production of voice and speech. Vocal biomarkers are signatures, features, or a combination of features from the audio signal of the voice, that is associated with a clinical outcome and can be used to monitor patients, diagnose a condition, or grade the severity of a disease. Vocal biomarkers for diseases affecting the nervous system are well-established, but there is also some evidence for a potential in diabetes and cardiovascular research. Therefore, this project focuses on cardiovascular disease (CVD), neuropathy, and diabetes distress as clinical outcomes. Previous studies have been rather small, therefore there is also a need to establish new data collection with a focus on diabetes-related complications.

Aims This interdisciplinary project aims to develop and integrate novel vocal biomarkers in risk assessment of diabetes-related complications. The work will involve (1) data collection, creating new resources for further research in an emerging field, and (2) development of machine learning methods and models that might reveal important clinical knowledge about diabetes-related complications: cardiovascular disease, neuropathy and diabetes distress.

Methods First, machine learning models will be pre-trained on large datasets (audio and image) for various audio prediction tasks. These models will then be fine-tuned for the clinical prediction tasks with a method called transfer learning. These models will predict the presence of CVD, neuropathy, and the level of diabetes distress. For prediction of diabetes distress, voice data will be combined with features extracted from answers to open-ended questions with large language models. Model performance will be evaluated in internal test sets and validated in a global datasource (Colive Voice) during a research stay at the Luxembourg Institute of Health.

Perspectives The proposed project will contribute with valuable insight on how voice data can be used in risk assessment of diabetes-related complications. The project is expected to generate both methodological results (e.g. pre-trained models, new data sources for machine learning research) and clinically relevant tools (e.g. vocal biomarkers) that might contribute to innovative ways of monitoring diabetes-related complications in the future.