Investigating the Use of Speech Analysis as a Diagnostic Tool for Treatment-Resistant Depression

Poster Session A - Saturday, April 13, 2024, 2:30 – 4:30 pm EDT, Sheraton Hall ABC

Micaela Wiseman^1,2 (micaela.wiseman@mail.utoronto.ca), Madeline Wood Alexander^1,3, Sean Nestor^1,4, Nir Lipsman^1,2,4,5, Jessica Robin⁷, Michael J. Spilka⁷, William Simpson^1,2,4, Jennifer S. Rabin^1,2,3,6; ¹Harquail Centre for Neuromodulation, Sunnybrook Research Institute, Toronto ON, ²Institute of Medical Sciences, University of Toronto, Toronto, ON, ³Rehabilitation Sciences Institute, University of Toronto, Toronto, ON, ⁴Department of Psychiatry, University of Toronto, Toronto, ON, ⁵Department of Surgery, University of Toronto, Toronto, ON, ⁶Division of Neurology, Department of Medicine, Sunnybrook Health Sciences Centre, University of Toronto, Toronto, ON, ⁷Winterlight Labs, Inc, Toronto, ON;

Background: Objective markers of depression are needed to enhance personalized clinical management. Here, we examined whether a validated speech assessment could detect speech differences between individuals with treatment-resistant depression (TRD) and healthy controls, and whether speech variables correlated with depression severity, as measured by the Hamilton Depression Rating Scale (HAMD). Methods: 50 patients with TRD (mean age=44.0±14.3, 44% female, mean HAMD=21.1±4.1) and 16 non-depressed participants (mean age=46.6±19.5, 81% female) completed a validated speech assessment developed by Winterlight Labs. Participants were asked: “Tell me how you’re feeling today.” Acoustic and linguistic features were computed with natural language processing libraries. Speech outcomes of interest included fundamental frequency (mean, variance), vocal intensity, pause duration, speech rate, and affective content of speech (valence, arousal, dominance). ANCOVAs assessed group differences and linear regression models assessed associations between speech characteristics and HAMD scores in patients. Analyses adjusted for age, sex, and English proficiency. Results: Compared to controls, individuals with TRD exhibited lower fundamental frequency (p=0.03, η2=0.07), longer pause duration (p=0.04, η2=0.07), slower speech rate (p=0.007, η2=0.12), and more emotionally arousing (p=0.007, η2=0.12) and negatively valenced language (p=0.02, η2=0.09). There were no group differences in fundamental frequency variability, sentiment dominance or vocal intensity. There were no significant associations between speech characteristics and HAMD scores. Conclusions: Several speech variables differed between individuals with TRD and controls but did not correlate with HAMD scores. As such, these speech metrics may capture phenotypic characteristics of TRD. Further research is needed to validate these findings and assess their clinical utility.

Topic Area: EMOTION & SOCIAL: Emotion-cognition interactions