Question-based computational language approach outperform ratings scale in discriminating between anxiety and depression
(2025) In Journal of Anxiety Disorders 112.- Abstract
Major Depression (MD) and General Anxiety Disorder (GAD) are the most common mental health disorders, which typically are assessed quantitatively by rating scales such as PHQ-9 and GAD-7. However, recent advances in natural language processing (NLP) and machine learning (ML) have opened up the possibility of question-based computational language assessment (QCLA). Here we investigate how accurate open-ended questions, using descriptive keywords or autobiographical narratives, can discriminate between participants that self-reported diagnosis of depression and anxiety, or health control. The results show that both language and rating scale measures can discriminate well, however, autobiographical narratives discriminate best between... (More)
Major Depression (MD) and General Anxiety Disorder (GAD) are the most common mental health disorders, which typically are assessed quantitatively by rating scales such as PHQ-9 and GAD-7. However, recent advances in natural language processing (NLP) and machine learning (ML) have opened up the possibility of question-based computational language assessment (QCLA). Here we investigate how accurate open-ended questions, using descriptive keywords or autobiographical narratives, can discriminate between participants that self-reported diagnosis of depression and anxiety, or health control. The results show that both language and rating scale measures can discriminate well, however, autobiographical narratives discriminate best between healthy and anxiety (ϕ = 1.58), as well as healthy and depression (ϕ = 1.38). Descriptive keywords, and to a certain extent autobiographical narratives, also discriminate better than summed scores of GAD-7 and PHQ-9 (ϕ=0.80 in discrimination between anxiety and depression), but not when individual items of these scales were analyzed by ML (ϕ=0.86 and ϕ=0.91 in item-level analysis of PHQ-9 and GAD-7, respectively). Combining the scales consistently elevated the discrimination even more (ϕ=1.39 in comparison between depression and anxiety), both in item-level and sum-scores analyses. These results indicate that QCLA measures often, but not in all cases, are better than standardized rating scales for assessment of depression and anxiety. Implication of these findings for mental health assessments are discussed.
(Less)
- author
- Tabesh, Mona
LU
; Mirström, Mariam
LU
; Böhme, Rebecca Astrid LU ; Lasota, Marta LU ; Javaherian, Yousef ; Agbotsoka-Guiter, Thibaud and Sikström, Sverker LU
- organization
- publishing date
- 2025-06
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- AI, Anxiety, Assessment, Depression, Language, Rating scales
- in
- Journal of Anxiety Disorders
- volume
- 112
- article number
- 103020
- publisher
- Elsevier
- external identifiers
-
- pmid:40279835
- scopus:105003227657
- ISSN
- 0887-6185
- DOI
- 10.1016/j.janxdis.2025.103020
- language
- English
- LU publication?
- yes
- id
- dcda5710-4721-49c5-8c37-cd67afca3b25
- date added to LUP
- 2025-08-01 09:58:46
- date last changed
- 2025-08-02 02:44:29
@article{dcda5710-4721-49c5-8c37-cd67afca3b25, abstract = {{<p>Major Depression (MD) and General Anxiety Disorder (GAD) are the most common mental health disorders, which typically are assessed quantitatively by rating scales such as PHQ-9 and GAD-7. However, recent advances in natural language processing (NLP) and machine learning (ML) have opened up the possibility of question-based computational language assessment (QCLA). Here we investigate how accurate open-ended questions, using descriptive keywords or autobiographical narratives, can discriminate between participants that self-reported diagnosis of depression and anxiety, or health control. The results show that both language and rating scale measures can discriminate well, however, autobiographical narratives discriminate best between healthy and anxiety (ϕ = 1.58), as well as healthy and depression (ϕ = 1.38). Descriptive keywords, and to a certain extent autobiographical narratives, also discriminate better than summed scores of GAD-7 and PHQ-9 (ϕ=0.80 in discrimination between anxiety and depression), but not when individual items of these scales were analyzed by ML (ϕ=0.86 and ϕ=0.91 in item-level analysis of PHQ-9 and GAD-7, respectively). Combining the scales consistently elevated the discrimination even more (ϕ=1.39 in comparison between depression and anxiety), both in item-level and sum-scores analyses. These results indicate that QCLA measures often, but not in all cases, are better than standardized rating scales for assessment of depression and anxiety. Implication of these findings for mental health assessments are discussed.</p>}}, author = {{Tabesh, Mona and Mirström, Mariam and Böhme, Rebecca Astrid and Lasota, Marta and Javaherian, Yousef and Agbotsoka-Guiter, Thibaud and Sikström, Sverker}}, issn = {{0887-6185}}, keywords = {{AI; Anxiety; Assessment; Depression; Language; Rating scales}}, language = {{eng}}, publisher = {{Elsevier}}, series = {{Journal of Anxiety Disorders}}, title = {{Question-based computational language approach outperform ratings scale in discriminating between anxiety and depression}}, url = {{http://dx.doi.org/10.1016/j.janxdis.2025.103020}}, doi = {{10.1016/j.janxdis.2025.103020}}, volume = {{112}}, year = {{2025}}, }