{"title":"Machine learning-driven development of a stratified CES-D screening system: optimizing depression assessment through adaptive item selection.","authors":"Ruo-Fei Xu, Zhen-Jing Liu, Shunan Ouyang, Qin Dong, Wen-Jing Yan, Dong-Wu Xu","doi":"10.1186/s12888-025-06693-8","DOIUrl":null,"url":null,"abstract":"<p><strong>Objective: </strong>To develop a stratified screening tool through machine learning approaches for the Center for Epidemiologic Studies Depression Scale (CES-D-20) while maintaining diagnostic accuracy, addressing the efficiency limitations in large-scale applications.</p><p><strong>Methods: </strong>Data were derived from the Chinese Psychological Health Guard Project (primary sample: n = 179,877; age 9-18) and China Labor-force Dynamics Survey (validation samples across age spans). We employed a two-stage machine learning approach: first applying Recursive Feature Elimination with multiple linear regression to identify core predictive items for total depression scores, followed by logistic regression for optimizing depression classification (CES-D ≥ 16). Model performance was systematically evaluated through discrimination (ROC analysis), calibration (Brier score), and clinical utility analyses (decision curve analysis), with additional validation using random forest and support vector machine algorithms across independent samples.</p><p><strong>Results: </strong>The resulting stratified screening system consists of an initial four-item rapid screening layer (encompassing emotional, cognitive, and interpersonal dimensions) for detecting probable depression (AUC = 0.982, sensitivity = 0.945, specificity = 0.926), followed by an enhanced assessment layer with five additional items. Together, these nine items enable accurate prediction of the full CES-D-20 total score (R<sup>2</sup> = 0.957). This stratified approach demonstrated robust generalizability across age groups (R<sup>2</sup> > 0.94, accuracy > 0.91) and time points. Calibration analyses and decision curve analyses confirmed optimal clinical utility, particularly in the critical risk threshold range (0.3-0.6).</p><p><strong>Conclusions: </strong>This study contributes to the refinement of CES-D by developing a machine learning-derived stratified screening version, offering an efficient and reliable approach that optimizes assessment burden while maintaining excellent psychometric properties. The stratified design makes it particularly valuable for large-scale mental health screening programs, enabling efficient risk stratification and targeted assessment allocation.</p>","PeriodicalId":9029,"journal":{"name":"BMC Psychiatry","volume":"25 1","pages":"286"},"PeriodicalIF":3.4000,"publicationDate":"2025-03-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11938587/pdf/","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"BMC Psychiatry","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1186/s12888-025-06693-8","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"PSYCHIATRY","Score":null,"Total":0}
引用次数: 0
Abstract
Objective: To develop a stratified screening tool through machine learning approaches for the Center for Epidemiologic Studies Depression Scale (CES-D-20) while maintaining diagnostic accuracy, addressing the efficiency limitations in large-scale applications.
Methods: Data were derived from the Chinese Psychological Health Guard Project (primary sample: n = 179,877; age 9-18) and China Labor-force Dynamics Survey (validation samples across age spans). We employed a two-stage machine learning approach: first applying Recursive Feature Elimination with multiple linear regression to identify core predictive items for total depression scores, followed by logistic regression for optimizing depression classification (CES-D ≥ 16). Model performance was systematically evaluated through discrimination (ROC analysis), calibration (Brier score), and clinical utility analyses (decision curve analysis), with additional validation using random forest and support vector machine algorithms across independent samples.
Results: The resulting stratified screening system consists of an initial four-item rapid screening layer (encompassing emotional, cognitive, and interpersonal dimensions) for detecting probable depression (AUC = 0.982, sensitivity = 0.945, specificity = 0.926), followed by an enhanced assessment layer with five additional items. Together, these nine items enable accurate prediction of the full CES-D-20 total score (R2 = 0.957). This stratified approach demonstrated robust generalizability across age groups (R2 > 0.94, accuracy > 0.91) and time points. Calibration analyses and decision curve analyses confirmed optimal clinical utility, particularly in the critical risk threshold range (0.3-0.6).
Conclusions: This study contributes to the refinement of CES-D by developing a machine learning-derived stratified screening version, offering an efficient and reliable approach that optimizes assessment burden while maintaining excellent psychometric properties. The stratified design makes it particularly valuable for large-scale mental health screening programs, enabling efficient risk stratification and targeted assessment allocation.
期刊介绍:
BMC Psychiatry is an open access, peer-reviewed journal that considers articles on all aspects of the prevention, diagnosis and management of psychiatric disorders, as well as related molecular genetics, pathophysiology, and epidemiology.