Depression screening using a non-verbal self-association task: A machine-learning based pilot study

Liu, Y., Song, Y., Lee, Naomi Anne ORCID: https://orcid.org/0000-0002-0973-6394, Bennett, D., Button, K., Greenshaw, A., Cao, B. and Sui, J. ORCID: https://orcid.org/0000-0002-4031-4456 (2022) Depression screening using a non-verbal self-association task: A machine-learning based pilot study. Journal of Affective Disorders, 310. pp. 87-95. ISSN Print ISSN: 0165-0327 Online ISSN: 1573-2517

	Text (Author Upload) depression_screening_ML_Liu_et_al_2022.pdf - Published Version Restricted to Repository staff only Download (780kB) | Request a copy
Preview	Text pilot study.pdf - Accepted Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) | Preview

Abstract

Background: Effective screening is important to combat the raising burden of depression and opens a critical time window for early intervention. Clinical use of non-verbal depression screening is nascent, yet a promising and viable candidate to supplement verbal screening. Differential self- and emotion-processing in depression patients were previously reported by non-verbal behavioural assessments, corroborated by neuroimaging findings of distinct neuroanatomical markers. Thus non-verbal validated brain-behaviour based self-emotion-related assessment data reflect physiological differences and may support individual level screening of depression. Methods: In this pilot study (n = 84) we collected two longitudinal sessions of behavioural assessment data in a laboratory setting. Depression was assessed using Beck Depression Inventory II (BDI-II), to explore optimal screening methods with machine-learning, and to establish the validity of adapting a novel behavioural assessment focusing on self and emotions for depression screening. Results: The best machine-learning model achieved high performance in depression screening, 10-Fold cross-validation (CV) Area Under the receiver operating characteristic Curve (AUC) of 0.90 and balanced accuracy of 0.81, using a Gradient Boosting algorithm. Prospective prediction using a model trained with session 1 data to predict session 2 depression status achieved a 10-Fold CV AUC of 0.77 and balanced accuracy of 0.66. We also identified interpretable behavioural signatures for depression patients based on the best model. Conclusion: The study supports the utility of using behavioural data as a viable and cost-effective solution for depression screening, with a potential wide range of applications in clinical settings.

Actions (login required)

View Item