Fine-tuning large language models for detecting negative self-referential language in smartphone text among youth with depression. Treves, I. N., Li, L. Y., Hod, L., Schwartz, A., Bloom, P. A., Spence, J., Rosenberg, A. S., Durham, K., Xu, X., Trivedi, E., Pagliaccio, D., Allen, N. B., Shankman, S. A., & Auerbach, R. P. Journal of Psychopathology and Clinical Science, American Psychological Association, US, 2026.
doi  abstract   bibtex   
Adolescent smartphone language provides a lens into negative self-referential thinking, which is central to major depressive disorder (MDD). Prior studies have linked language features, including negative sentiment and first-person pronouns, to mood and depression, suggesting that naturalistic language may identify who is at risk and when that risk is greatest. However, studies of adolescent smartphone social communication have typically relied on rule-based models not validated for heterogeneous, context-sensitive language. To address this gap, we determined whether transformer models, including large language models, optimized detection of depression risk in extensive adolescent smartphone text data. In this study, 223 adolescents (Mage = 16.43 years, current MDD = 37, remitted MDD = 103, healthy controls = 83) installed a smartphone app, which prompted participants to provide mood ratings once a day and acquired all keyboard inputs over 12 months (mean text entries per participant = 17,683). Ten thousand text entries were double-coded for entry-level sentiment (positive, neutral, negative) and self-reference for training and testing traditional rule-based approaches (VADER, pronoun counts) and transformer approaches, including GPT-4-mini. In the full data set, between-participant associations with depressive symptoms and within-participant relations to daily mood and depressive episodes were examined. Fine-tuned transformer models best aligned with human-coded sentiment labels (F1GPT4-MINI = .84, F1VADER = .59) and accurately detected self-reference (F1T5-BASE = .97). Adolescents with current and remitted MDD exhibited more transformer-based negative self-referential language than healthy controls (ORs = 1.38, 1.26). Increased negative self-referential language predicted worse next-day mood (β = −.033, p \textless .001) and a higher likelihood of next-week depressive episodes (OR = 1.66, 95% confidence interval [1.06, 2.62], p = .028), although the association with depressive episodes was not robust to sensitivity analyses. Transformer models may be integrated into digital mental health care to detect when youth are at risk for depression. (PsycInfo Database Record (c) 2026 APA, all rights reserved)
@article{treves_fine-tuning_2026,
	address = {US},
	title = {Fine-tuning large language models for detecting negative self-referential language in smartphone text among youth with depression},
	copyright = {All rights reserved},
	issn = {2769-755X},
	doi = {10.1037/abn0001159},
	abstract = {Adolescent smartphone language provides a lens into negative self-referential thinking, which is central to major depressive disorder (MDD). Prior studies have linked language features, including negative sentiment and first-person pronouns, to mood and depression, suggesting that naturalistic language may identify who is at risk and when that risk is greatest. However, studies of adolescent smartphone social communication have typically relied on rule-based models not validated for heterogeneous, context-sensitive language. To address this gap, we determined whether transformer models, including large language models, optimized detection of depression risk in extensive adolescent smartphone text data. In this study, 223 adolescents (Mage = 16.43 years, current MDD = 37, remitted MDD = 103, healthy controls = 83) installed a smartphone app, which prompted participants to provide mood ratings once a day and acquired all keyboard inputs over 12 months (mean text entries per participant = 17,683). Ten thousand text entries were double-coded for entry-level sentiment (positive, neutral, negative) and self-reference for training and testing traditional rule-based approaches (VADER, pronoun counts) and transformer approaches, including GPT-4-mini. In the full data set, between-participant associations with depressive symptoms and within-participant relations to daily mood and depressive episodes were examined. Fine-tuned transformer models best aligned with human-coded sentiment labels (F1GPT4-MINI = .84, F1VADER = .59) and accurately detected self-reference (F1T5-BASE = .97). Adolescents with current and remitted MDD exhibited more transformer-based negative self-referential language than healthy controls (ORs = 1.38, 1.26). Increased negative self-referential language predicted worse next-day mood (β = −.033, p {\textless} .001) and a higher likelihood of next-week depressive episodes (OR = 1.66, 95\% confidence interval [1.06, 2.62], p = .028), although the association with depressive episodes was not robust to sensitivity analyses. Transformer models may be integrated into digital mental health care to detect when youth are at risk for depression. (PsycInfo Database Record (c) 2026 APA, all rights reserved)},
	journal = {Journal of Psychopathology and Clinical Science},
	publisher = {American Psychological Association},
	author = {Treves, Isaac N. and Li, Lilian Y. and Hod, Libby and Schwartz, Adela and Bloom, Paul A. and Spence, Jamaal and Rosenberg, Alexander Saul and Durham, Katherine and Xu, Xuhai and Trivedi, Esha and Pagliaccio, David and Allen, Nicholas B. and Shankman, Stewart A. and Auerbach, Randy P.},
	year = {2026},
	keywords = {Emotional States, Smartphones, Major Depression, Adolescent Psychopathology, Text Messaging, Large Language Models, Pronouns, Self-Reference},
	file = {Full Text PDF:/Users/mexico/Zotero/storage/SWMYNUWC/Treves et al. - 2026 - Fine-tuning large language models for detecting negative self-referential language in smartphone tex.pdf:application/pdf},
}
Downloads: 0
About
Documentation
Keywords
Search
Users
Terms and Conditions of Use
Privacy Policy
Sitemap
© BibBase.org 2021