Demographics and text predictors of PHQ-2-measured depression using machine learning

Background The Patient Health Questionnaire-2 (PHQ-2) is a two-item screener widely used in large population surveys. When brief instruments are used as predictive outcomes, the information available for discrimination may be limited. Methods We analyzed repeated cross-sectional data from the 2021–2024 Household Pulse Survey, including 605,877 adults after complete-case restrictions. Depression was defined dichotomously as PHQ-2 score greater than or equal to three. Random Forest, Extreme Gradient Boosting, and logistic regression were trained using demographic predictors, Term Frequency-Inverse Document Frequency (TF-IDF)-based feature representations derived from demographic variables, and a combined feature set. Model performance was evaluated using cross-validated discrimination metrics and calibration assessment. PHQ-2 reliability was estimated using McDonald’s omega and split-half reliability (i.e., Pearson correlation with Spearman Brown formula). Results PHQ-2 reliability was high across survey waves (ω = 0.86–0.87); split-half Spearman–Brown coefficients ranged from 0.75 to 0.77. Model discrimination was similar across both machine learning methods and logistic regression, as well as across predictor sets (AUC: 0.68–0.69). Although Extreme Gradient Boosting performed slightly better than Random Forest, logistic regression demonstrated comparable performance, and differences across predictor sets (demographics, TF-IDF features, and combined features) were minimal within each modeling approach, indicating that model performance was not strongly influenced by either model complexity or the choice of input features. Conclusions Across modeling approaches and feature sets, predictive performance was comparable and modest. When a two-item, dichotomized screener is used as the outcome, gains from alternative feature representations or increased model complexity are limited by either the set of predictors or the informational content of the ultra-brief depression measure.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-19
DOI
https://doi.org/10.1371/journal.pone.0361381
Primary Topic
Artificial Intelligence in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Demographics and text predictors of PHQ-2-measured depression using machine learning

Patrick O. Monahan, Kurt Kroenke, Felix M. Pabon‐Rodriguez, Nathaniel J. Maxey et al.
PLoS ONE
Artificial Intelligence in Healthcare
article

Demographics and text predictors of PHQ-2-measured depression using machine learning

Patrick O. Monahan, Kurt Kroenke, Felix M. Pabon‐Rodriguez, Nathaniel J. Maxey, Brian Spitzer
article en

Abstract

Background The Patient Health Questionnaire-2 (PHQ-2) is a two-item screener widely used in large population surveys. When brief instruments are used as predictive outcomes, the information available for discrimination may be limited. Methods We analyzed repeated cross-sectional data from the 2021–2024 Household Pulse Survey, including 605,877 adults after complete-case restrictions. Depression was defined dichotomously as PHQ-2 score greater than or equal to three. Random Forest, Extreme Gradient Boosting, and logistic regression were trained using demographic predictors, Term Frequency-Inverse Document Frequency (TF-IDF)-based feature representations derived from demographic variables, and a combined feature set. Model performance was evaluated using cross-validated discrimination metrics and calibration assessment. PHQ-2 reliability was estimated using McDonald’s omega and split-half reliability (i.e., Pearson correlation with Spearman Brown formula). Results PHQ-2 reliability was high across survey waves (ω = 0.86–0.87); split-half Spearman–Brown coefficients ranged from 0.75 to 0.77. Model discrimination was similar across both machine learning methods and logistic regression, as well as across predictor sets (AUC: 0.68–0.69). Although Extreme Gradient Boosting performed slightly better than Random Forest, logistic regression demonstrated comparable performance, and differences across predictor sets (demographics, TF-IDF features, and combined features) were minimal within each modeling approach, indicating that model performance was not strongly influenced by either model complexity or the choice of input features. Conclusions Across modeling approaches and feature sets, predictive performance was comparable and modest. When a two-item, dichotomized screener is used as the outcome, gains from alternative feature representations or increased model complexity are limited by either the set of predictors or the informational content of the ultra-brief depression measure.

PLoS ONEVol. 21(11)
Regenstrief Institute (US), Indiana University Indianapolis (US), Indiana University School of Medicine (US), New York University (US)
Artificial Intelligence in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon