Associate Professor Visar Berisha, Electrical, Computer, and Energy Engineering and College of Health Solutions, Arizona State U., USA.
- Title: Developing speech-based clinical machine learning models that work: should we believe reported accuracies in the academic literature?
- Abstract: The most common approach to speech-based clinical machine learning models is supervised learning. That is, the machine learning (ML) algorithm designer collects speech samples and clinical labels from clinical groups of interest, then they train an ML model to predict the label from features extracted from the speech, and report the accuracy of the model. We analyze 77 published studies reporting on the accuracy of speech-based models for predicting dementia, combined across three different meta-analyses, and provide evidence of overoptimism in the external validity of these models. We provide possible explanations for the reasons behind this overoptimism in the literature and discuss several strategies for building robust clinical machine learning models that generalize to real-world conditions.
Dr. Athanasios Tsanas, the Usher Institute, University of Edinburgh, UK.
- Title: Harnessing voice signals using signal processing and statistical machine learning: applications in mental health and other biomedical and life sciences applications.
- Abstract: I will briefly outline the main physiological principles of voice production and describe how these link to the key concepts for developing speech signal processing algorithms to characterize speech and extract potentially useful information. I will demonstrate the applicability and differences of speech signal processing algorithmic concepts across different applications, in combination with state of the art statistical machine learning techniques. Finally, I will touch on open questions, challenges, and upcoming problems as we develop robust, parsimonious, generalizable decision support tools mining speech signals across diverse biomedical and life sciences applications.