Luxembourg Institute of Health (LIH)
Abstract: Depression affects around 4.4% of the global population, requiring early and accurate screening to reduce its long-term damage. Traditional screening approaches such as self-reported questionnaires have limitations, prompting the shift to objective assessment methods such as voice-based biomarkers. These biomarkers offer non-invasive and scalable depression screening, though current models demonstrate only modest accuracy. An underexplored avenue is the correlation between other symptoms, like fatigue and sleeping disorders, and their potential to improve the efficacy of vocal biomarkers. We anticipate performance enhancement by integrating multitask learning, considering additional symptoms. Multitask learning, grounded in its ability to address complex problems with associated correlations, can refine generalization through domain information from interconnected tasks. We will use a large dataset from the Colive Voice study, which has voice samples from over 6,000 participants in multiple languages. The samples include voice tasks like reading text, counting numbers, and /a/ phonation. Along with these voice tasks, we’re looking at questionnaire data on depression, fatigue, and sleep disorders. Multitask learning will simultaneously teach a computer model to recognize signs of all three conditions. By doing this, we believe the model will become better at identifying each symptom by understanding how they relate to each other.
Keywords: Depression screening, vocal biomarkers, multitask learning, deep learning, convolutional neural networks, digital health
Scientific area: Digital Health and Artificial Intelligence
Visiting period: 2 weeks
Visiting Lab: German Research Center for Artificial Intelligence (DFKI)
Bio: Abir is a PhD student in the Deep Digital Phenotyping research unit at LIH. Her research focus is on the identification of vocal biomarkers to monitor health in people with serious/chronic illnesses with the use of artificial intelligence methods.