Assessing Depression Severity From Speech: The Role of Clinical Intuition and Artificial Intelligence

Abstract

Automated speech analysis methods are used to estimate depression severity. However, it is unclear how these compare to the intuition of expert clinicians using the same information from speech. We distributed 1-minute speech recordings from 72 participants to 12 clinicians, who estimated depression severity after listening to speech recordings. We trained acoustic and text-based AI models to estimate depression severity in the same samples. Clinicians had a higher agreement to MADRS scores (ICC= 0.47, 95%) than the acoustic-based (ICC = 0.35), text-based (ICC = 0.29), and combined acoustic and text (ICC = 0.33) AI model estimations. However, clinicians had larger errors (RMSE = 10.98) than the text-based (RMSE = 10.02), acoustic-based (RMSE = 10.69), and combined (RMSE = 9.71) AI models. Bias analysis showed clinician gender-based differences in depression estimation. These findings provide the first direct comparison of clinical intuition and AI estimation of depression severity from speech.