The research team, led by Shu-Fen Cheng, Yun-Ying Ding, and corresponding author Hua-Dong Hong, drew their evidence from the China Health and Retirement Longitudinal Study, known as CHARLS, a nationally representative survey that follows thousands of Chinese households over time. From the 2018 wave, the investigators identified 1,269 patients with gastrointestinal diseases who were at least 45 years old. Depression was defined using the ten-item Center for Epidemiologic Studies Depression Scale, or CES-D-10, a widely used brief questionnaire in which a score of 10 or more signals clinically relevant depressive symptoms. By this threshold, nearly half of the study population — a baseline prevalence of 49.2 percent — screened positive, a figure that underscores just how intertwined gut disease and mood disturbance become in later life.

What makes the study methodologically interesting is the way the team chose which variables to keep. Rather than dumping every available survey item into a model and hoping for the best, they used a hybrid feature-selection strategy combining Boruta, an algorithm built around randomized decision forests that tests each variable against shuffled copies of itself, with LASSO regression, a technique that shrinks the coefficients of weak predictors all the way to zero. The result was a parsimonious set of six core predictors: self-rated health status, trouble with body pain, gender, difficulty with instrumental activities of daily living such as shopping and managing medications, sleep duration, and limitations in basic activities of daily living like dressing and bathing. Six variables is a number a nurse could collect in a couple of minutes.

Then came the contest. The researchers pitted six machine learning algorithms against one another — logistic regression, K-nearest neighbors, support vector machine, random forest, gradient boosting machine, and extreme gradient boosting — tuning each model’s hyperparameters through ten-fold cross-validation, a procedure that repeatedly splits the data so that every case gets a turn in the test set. The verdict was, on its face, a small embarrassment for the fancy algorithms: plain, old-fashioned logistic regression, a statistical method older than most of the people studying it, showed the best generalizability on the validation cohort. Its area under the receiver operating characteristic curve, the standard measure of discrimination between those with and without the condition, reached 0.697, with a 95 percent confidence interval of 0.644 to 0.748.

Those numbers describe a model of moderate, not spectacular, power, and the authors are candid about that. The logistic regression model achieved an accuracy of 0.651, a sensitivity of 0.722, a specificity of 0.583, and an F1-score of 0.670. In practical terms, the tool catches roughly seven in ten people who are depressed, while correctly clearing a bit more than half of those who are not. For a screening instrument — as opposed to a diagnostic test — that trade-off is defensible, because the cost of a false alarm is merely a fuller clinical evaluation, whereas the cost of a missed case can be months or years of untreated illness. Notably, the area under the precision-recall curve, 0.671, exceeded the baseline prevalence of 0.492, indicating the model adds genuine information beyond simply guessing from the base rate.

Discrimination is only half the story, however. A screening model also needs to produce probabilities that mean what they say. The team assessed calibration using the Brier score, which came in at 0.225, and the Hosmer-Lemeshow test, which returned a p-value of 0.614 — a result that, unusually, is good news, since a high p-value here indicates no statistically significant departure between predicted and observed risk. Decision curve analysis, a technique that quantifies the net clinical benefit of acting on a model’s predictions across a range of risk thresholds, showed positive net benefit across threshold probabilities from 20 to 70 percent. In other words, across most of the range where a clinician might plausibly decide to intervene, using the model beats the two crude alternatives of screening everyone or screening no one.

Perhaps the most visually compelling contribution is the study’s use of SHAP, or SHapley Additive exPlanations, a framework borrowed from cooperative game theory that assigns each feature a fair share of credit for every individual prediction. Rather than treating the model as a black box, SHAP lets researchers and clinicians see exactly how much each variable pushes a given patient’s risk up or down. In this study, SHAP ranked the six predictors by contribution in descending order: health status first, then trouble with body pain, gender, instrumental activities of daily living, sleep time, and activities of daily living. The ordering tells a coherent clinical story — how patients feel overall, and how much pain interferes with their lives, matters more for depression risk than demographic or functional details alone.

The final product is a nomogram, a graphical scoring device that converts each of the six variables into points on a scale, sums them, and translates the total into an estimated probability of depressive symptoms. Nomograms have a long pedigree in medicine, particularly in oncology, precisely because they demand nothing more exotic than a pencil and a printed chart. In settings without reliable internet access or electronic health record integration — which describes a great many clinics where older gastrointestinal patients actually receive care — that simplicity is not a limitation but the entire point. The authors position the tool as a supplementary screening aid for the concurrent identification of depressive symptoms, enabling timely integrated care rather than replacing formal psychiatric assessment.

The study’s limitations deserve honest airing, and the authors supply them. This was a cross-sectional analysis, meaning it captures a single snapshot in time and cannot establish that the measured factors cause depression, only that they travel with it. The validation was internal, performed through resampling within the same dataset rather than on an independent external cohort, and the authors state plainly that external validation is urgently needed before the tool can be recommended broadly. The CHARLS sample, while nationally representative of China, may not generalize to other populations, and self-reported gastrointestinal disease and survey-based measures carry their own measurement error. An AUC just under 0.70 leaves meaningful room for improvement, and no screening chart should ever override a clinician’s judgment or a patient’s own account of their suffering.

Still, the study’s broader message resonates well beyond its specific numbers. At a moment when machine learning is often sold as a substitute for clinical reasoning, this work demonstrates a humbler and arguably more useful paradigm: use the sophisticated algorithms to interrogate the data, then deploy the simplest model that does the job, wrapped in transparency tools that let clinicians see why it predicts what it predicts. For the millions of middle-aged and older adults whose gut ailments and low moods feed each other in silence, a six-question chart that takes minutes to complete could be the difference between depression noticed early and depression discovered late. The research was supported by the Suzhou Science and Technology Development Program and the Suzhou Medical Association, and the underlying CHARLS data were approved by the Ethical Review Committee of Peking University, with all participants providing informed consent. The article is published open access, making the full methods and findings freely available to clinicians and researchers worldwide.

Subject of Research: Depression risk screening in middle-aged and older patients with gastrointestinal diseases using a logistic-regression nomogram

Article Title: A parsimonious logistic-regression nomogram for depression risk screening in middle-aged and older patients with gastrointestinal diseases: a cross-sectional CHARLS study with ML-Style comparison and SHAP visualization

Article References: Cheng, S.-F., Ding, Y.-Y., & Hong, H.-D. (2026). A parsimonious logistic-regression nomogram for depression risk screening in middle-aged and older patients with gastrointestinal diseases: a cross-sectional CHARLS study with ML-Style comparison and SHAP visualization. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08731-5

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08731-5

Keywords: depression, gastrointestinal diseases, CHARLS, logistic regression, nomogram, SHAP, machine learning, screening, CES-D-10, older adults, risk prediction, BMC Psychiatry