Friday, November 20, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Psychology & Psychiatry

Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study

November 12, 2026
in Psychology & Psychiatry
Glenn Wilkins
By Glenn Wilkins Scienmag Editorial Profile - Clinical Psychology
Reading Time: 6 mins read
0
Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study

Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

A computational screening framework has demonstrated that the standard ninety-item Symptom Checklist-90 questionnaire can be trimmed to forty-four questions while preserving classification agreement with the full-length assessment, according to an evaluation conducted with vocational college students in China. The investigation, undertaken by Meng Li, Cheng Liu, and Gang Shao at the Jiangsu Vocational Institute of Commerce, sought to resolve long-standing practical difficulties linked to the widely deployed mental health inventory. Although the Symptom Checklist-90 has functioned as an institutional benchmark across psychiatric and educational environments for nearly three decades, its substantial length routinely induces participant fatigue, elevating the proportion of careless responses and undermining diagnostic screening fidelity. By pairing gradient boosting classifiers with cooperative game theory metrics, the investigators developed a parsimonious forty-four-item proxy that captured the primary screening labels generated by the complete instrument, offering a proof-of-concept pathway toward reducing respondent burden in resource-limited academic settings.

The underlying investigation evaluated the responses of 5,126 vocational college students drawn from an initial stratified random sample of 5,420 returned online questionnaires administered between September 2023 and March 2024. Questionnaires completed in less than five minutes or containing direct logical contradictions were discarded during data cleaning. The final analyzed cohort comprised 2,461 male participants and 2,665 female participants across three academic cohorts, consisting of 1,632 freshmen, 1,980 sophomores, and 1,514 juniors. In standard Chinese screening protocols, respondents are operationalized as screening positive for elevated psychological distress if their total Symptom Checklist-90 score exceeds 160 or if they endorse more than 43 individual items with a symptom severity rating of two or higher on the five-point Likert scale. Within this vocational student population, 764 participants, representing 14.9 percent of the sample, met the operational criteria for mental health screening positivity, whereas 4,362 participants exhibited no significant screening risks. Overall questionnaire scores averaged 117.40 with a standard deviation of 42.79, reflecting mild average symptom manifestations across all nine evaluated psychological dimensions.

To establish baseline predictive capacity across the full inventory, the researchers first benchmarked nine distinct classification algorithms using ten-fold cross-validation on the entire ninety-item feature set. The tested algorithms included Logistic Regression, Decision Tree, Extra Trees, Random Forest, Multi-layer Perceptron, Support Vector Classifier, Gradient Boosting Machine, Light Gradient Boosting Machine, and Extreme Gradient Boosting. Light Gradient Boosting Machine attained the highest overall accuracy at 0.9830 and the leading baseline F1 score of 0.9413, supported by a precision of 0.9620 and a recall of 0.9221. The baseline Extreme Gradient Boosting model recorded an F1 score of 0.9348, while Random Forest achieved the peak precision of 0.9860 alongside a reduced recall of 0.8604. Support Vector Classifier demonstrated the highest recall at 0.9411. Because the dataset exhibited marked class imbalance, with positive cases accounting for less than fifteen percent of all observations, the research team designated the F1 score as the primary operational metric to penalize models that inflated overall accuracy through majority-class bias.

To eliminate redundant questions without compromising predictive reliability, the investigators introduced a feature selection strategy governed by SHapley Additive exPlanations, commonly designated as SHAP. Grounded in cooperative game theory, the TreeSHAP algorithm calculated the marginal contribution of each symptom query toward the overall classification output across default tree models. Items were arranged in descending order based on their mean absolute SHAP values, after which sequential forward selection iteratively expanded the feature subset. For each newly formed subset, the team trained gradient boosting architectures and monitored classification trajectories across ten-fold cross-validation. The stopping threshold was governed by a pre-specified one-standard-error rule, which designated the optimal configuration as the smallest item set whose cross-validated F1 score remained within one standard error of the trajectory peak. This experimental stopping rule isolated 44 items for Extreme Gradient Boosting, 59 items for Light Gradient Boosting Machine, and 38 items for the standard Gradient Boosting Machine, with the 44-item Extreme Gradient Boosting model achieving the highest comparative F1 score of 0.9471.

Following feature reduction, the researchers tested four bio-inspired metaheuristic optimization algorithms to refine model hyperparameters: Grey Wolf Optimization, Sparrow Search Algorithm, Whale Optimization Algorithm, and Particle Swarm Optimization. The Grey Wolf Optimizer, which mimics pack hunting mechanisms including encirclement and directed pursuit, yielded the strongest hyperparameter combinations for the 44-item Extreme Gradient Boosting classifier across ten repeated experimental iterations. Under the initial full-data evaluation protocol, the tuned 44-item pipeline achieved an out-of-fold F1 score of 0.9501, outperforming random selection at 0.9230, LASSO regression selection at 0.9361, and mutual information selection at 0.9270 matched to the identical 44-item budget. However, because feature selection and tuning conducted on the aggregate dataset introduce optimistic selection bias, the authors subsequently re-evaluated the pipeline under a strict, leakage-free nested ten-fold cross-validation design where item ranking, stopping rules, and hyperparameter searches were confined exclusively to the inner training folds.

Under the leakage-free nested validation protocol, the tuned 44-item Extreme Gradient Boosting model achieved an F1 score of 0.9233 with a standard deviation of 0.0248 and an Area Under the Precision-Recall Curve of 0.9853. Statistical testing via the corrected resampled t-test indicated that this performance did not differ significantly from the untuned full-feature baseline score of 0.9405 (p = 0.188), a conclusion reaffirmed by a non-parametric Wilcoxon signed-rank test (p = 0.074). When evaluated on a frozen, stratified hold-out validation cohort of 1,026 participants partitioned before model construction, the frozen 44-item pipeline correctly identified 138 of 153 positive screening cases and 869 of 873 negative cases. This single hold-out test demonstrated an accuracy of 0.9815, a precision of 0.9718, a recall of 0.9020, an F1 score of 0.9356, an Area Under the Precision-Recall Curve of 0.9883, and an exemplary Brier calibration score of 0.015, while decision-curve analysis confirmed sustained clinical net benefit across threshold probabilities spanning 0.01 to 0.80.

The study also reconstructed four historically published Symptom Checklist-90 short forms using the same student response dataset to conduct head-to-head empirical comparisons against the machine learning pipeline. In standard ten-fold evaluations, the 27-item SCL-27 reached an F1 score of 0.9011, the 18-item Brief Symptom Inventory reached 0.8765, the 14-item SCL-14 achieved 0.8685, and the 9-item SCL-K-9 recorded 0.8589. While these shorter forms are inherently unaffected by feature-selection optimism due to their fixed item compositions, all four fell substantially below the 0.9233 nested estimate attained by the optimized 44-item subset. Resampling stability analysis incorporating 200 bootstrap iterations revealed that the 44 selected items maintained a median selection frequency of 94 percent across resamples, yielding an average Jaccard similarity index of 0.71, which confirmed the presence of a resilient core symptom nucleus despite localized permutations among correlated variables.

Interpretability analysis conducted on the 44-item model provided granular insights into how individual psychological items drove screening predictions. The five most influential individual queries identified by SHAP importance were item S32 measuring loss of interest in things, item S30 capturing feeling blue, item S71 assessing the sensation that everything requires effort, item S90 reflecting the intrusive belief that something is wrong with one’s mind, and item S34 denoting having one’s feelings easily hurt. When examining composite dimensions, depression emerged as the dominant predictive factor with a mean absolute SHAP value of 2.199, followed by psychoticism at 1.901 and anxiety at 1.898. Interestingly, paranoid ideation advanced from ninth place in the full-scale baseline model to sixth place in the reduced model. Locally weighted scatterplot smoothing regression revealed that somatization exhibited an inflection point at a score of 1.67, where its marginal contribution to predicted screening risk transitioned from protective or neutral to positive.

Analysis of inter-factor interactions identified the pairing of depression and anxiety as the largest synergistic effect in the model, recording a mean SHAP interaction intensity of 0.322, followed closely by the interaction between psychoticism and depression at 0.317. The authors specifically noted that these synergistic values represent descriptive statistical properties of the non-linear machine learning model fitted to this specific cohort, rather than direct proof of clinical symptom reinforcement in the broader population. The elevated relevance of somatization within the predictive rankings also mirrored broader cultural patterns observed in Chinese vocational student cohorts, where distress frequently manifests through physical complaints rather than direct affective disclosure, contrasting sharply with Western student datasets where somatization rarely features among top predictive screening indices.

Despite the high predictive accuracy and substantial questionnaire reduction, the authors cautioned that the 44-item inventory cannot currently serve as an independent diagnostic replacement for the original scale. Although the total 44-item score correlated near-perfectly with the full 90-item total score at r > 0.99, confirmatory factor analysis on the 42 dimensional items revealed marginal structural fit indices: a comparative fit index of 0.879, a Tucker-Lewis index of 0.867, a root mean square error of approximation of 0.074, and a standardized root mean square residual of 0.130. Because these psychometric metrics fall short of conventional structural validity thresholds, the authors stressed that the condensed questionnaire must be treated as a proof-of-concept screening proxy for reproducing classification status, pending formal multi-center external validation and longitudinal reliability assessments.

Subject of Research: Optimization of the Symptom Checklist-90 psychiatric screening tool using machine learning and SHAP feature selection in vocational college students

Article Title: Optimization of the SCL-90 scale using machine learning: a novel approach to mental health assessment

Article References: Li, M., Liu, C., & Shao, G. (2026). Optimization of the SCL-90 scale using machine learning: a novel approach to mental health assessment. Frontiers in Psychology, 17, Article 2018014. https://doi.org/10.3389/fpsyg.2026.2018014

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.2018014

Keywords: Symptom Checklist-90, psychiatric screening, machine learning, SHapley Additive exPlanations, TreeSHAP, XGBoost, LightGBM, mental health assessment, Grey Wolf Optimization, college student mental health, vocational college students, nested cross-validation, somatization, depression screening, psychometrics, feature selection, confirmatory factor analysis

Cite Scienmag News

Glenn Wilkins. (November 12, 2026). Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study. Scienmag. https://scienmag.com/machine-learning-compresses-scl-90-questionnaire-by-half-in-college-screening-study/

Glenn Wilkins. "Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study." Scienmag, 12 November 2026, https://scienmag.com/machine-learning-compresses-scl-90-questionnaire-by-half-in-college-screening-study/. Accessed 20 November 2026.

Glenn Wilkins. "Machine Learning Compresses SCL-90 Questionnaire by Half in College Screening Study." Scienmag. November 12, 2026. https://scienmag.com/machine-learning-compresses-scl-90-questionnaire-by-half-in-college-screening-study/

Tags: college student mental healthconfirmatory factor analysisdepression screeningfeature selectiongrey wolf optimizationLightGBMMachine learningmental health assessmentnested cross-validationpsychiatric screeningpsychometricsShapley additive explanationssomatizationsymptom checklist-90TreeSHAPvocational college studentsXGBoost
Share26Tweet16
Previous Post

Egyptian Study Finds Higher MRSA Prevalence in Healthcare Workers

Next Post

Lebanese Survey Finds Most Adult Respondents Use AI Chatbots for Psychological Support

Related Posts

Restorative Campus Environments Link to Higher Study Engagement in College Survey
Psychology & Psychiatry

Restorative Campus Environments Link to Higher Study Engagement in College Survey

November 20, 2026
Parental Emotion Regulation Linked to Preschoolers’ Social Competence in China
Psychology & Psychiatry

Parental Emotion Regulation Linked to Preschoolers’ Social Competence in China

November 20, 2026
Disruptions in International Assistance Threaten Decades of Infectious Disease Progress
Psychology & Psychiatry

Disruptions in International Assistance Threaten Decades of Infectious Disease Progress

November 20, 2026
Qualitative Study Maps Diverse Self-Identities and Roles Across BDSM and Kink Sexuality
Psychology & Psychiatry

Qualitative Study Maps Diverse Self-Identities and Roles Across BDSM and Kink Sexuality

November 20, 2026
Eco-Friendly Additions Can Make High-Impact Shopping Sequences Seem Worse
Psychology & Psychiatry

Eco-Friendly Additions Can Make High-Impact Shopping Sequences Seem Worse

November 20, 2026
Meeting Postpartum 24-Hour Movement Guidelines Linked to Lower Depression and Stress
Psychology & Psychiatry

Meeting Postpartum 24-Hour Movement Guidelines Linked to Lower Depression and Stress

November 20, 2026
Next Post
Lebanese Survey Finds Most Adult Respondents Use AI Chatbots for Psychological Support

Lebanese Survey Finds Most Adult Respondents Use AI Chatbots for Psychological Support

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Dental Therapy Under Anesthesia Prepares Pediatric Transplant Patients in Case Series
  • Study Finds Documentation Gaps in Triple-Negative Breast Cancer Mastectomy Care
  • German Workers Favor Non-EU Skilled Immigrants with Job Offers and Qualifications
  • Genomize-SEQ Platform Accurately Identifies Rare Disease Variants in Exome Sequencing Benchmark

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading