Voice- and speech-based markers of neuropsychiatric conditions

Voice- and speech-based markers of neuropsychiatric conditions: assessing methodological and ethical foundations for clinical application
Abstract
Promising ML applications have shown great potential to identify vocal and speech markers of the most important neuropsychiatric conditions (e.g., Hitczenko et al., 2021; Cohen et al., 2021; Corcoran et al., 2020) and to develop systems able to monitor patients' symptoms and assist clinicians during the assessment process. However, these efforts face important limitations: the limited replicability and generalizability of previous results (Parola et al., 2022; Fusaroli et al., 2021), few attempts to explicitly account for the heterogeneity of the disorders (Mittal, 2021), and no concrete translation into effective clinical applications yet. What is critically lacking is an explicit reflection on the risks and limitations of ML applications that can support the development of robust, effective, and ethically grounded translational work. In this project, we will explore avenues to improve the clinical impact of ML applications in speech and voice research, focusing primarily on applicability and ethical concerns.
To this end, we draw on an already collected large dataset of voice and speech samples from the Danish High Risk and Resilience Study - VIA7 study (Gantriis et al., 2019), which examined 522 children born to parents diagnosed with schizophrenia (SZ) or bipolar disorder (BP). Our goal is to develop conservative (i.e., more robust and generalizable) ML and NLP pipelines to identify vocal and language markers of clinical symptoms in children at high-risk, that can serve as a reference for future studies. In addition, we aim to assess the impact of heterogeneity (e.g., socioeconomic, demographic, and clinical differences) and the presence of potential methodological biases and limitations, and robustly test the reliability of the results against various preprocessing and analytical procedure. Finally, we will explore how ML techniques can concretely support the development of robust, effective, and ethically founded clinical applications, and evaluate how to include from the very design of a study a consideration of risks, limitations, and ethical practices. The final outcome is to provide a first solid effort - both conceptually and methodologically - for the development of better practices in ML, SSP and NLP clinical research.
About the speaker
Alberto Parola, postdoc, Department of Linguistics, Cognitive Science and Semiotics, Aarhus University
IMC Tuesday Seminar held June 20th, 2023.
Note: Talk is trimmed to ensure anonymity of informants.