Simple Logistic Regression Shows Stable Screening Performance for Depressive Symptoms in COPD

A newly published investigation in BMC Pulmonary Medicine demonstrates that an interpretable logistic regression model can identify community-dwelling individuals with chronic obstructive pulmonary disease who are at high risk of depressive symptoms. By systematically evaluating multiple supervised machine learning architectures, the investigators established that a parsimonious model relying on only six routinely collectible predictors achieved the most reliable and stable generalization performance.

Chronic obstructive pulmonary disease is frequently accompanied by depressive symptoms. However, these depressive symptoms are often under-recognized in community settings. Although predictive algorithms employing artificial intelligence and machine learning have recently emerged to estimate depression risk among pulmonary patients, empirical evidence directly derived from community-dwelling Chinese populations has remained notably scarce. The primary objective of the researchers was therefore to systematically compare commonly utilized supervised machine learning algorithms to construct a parsimonious, clinically interpretable risk prediction framework capable of early detection among community-based patients.

To construct and rigorously validate the predictive models, the authors examined an epidemiological cohort drawn from the 2020 wave of the China Health and Retirement Longitudinal Study, widely known as CHARLS. The final analytical sample comprised 809 community-dwelling participants aged 40 years or older who reported a physician diagnosis of chronic obstructive pulmonary disease or related chronic lung disease. In defining the primary psychological outcome, the investigators utilized the validated 10-item Center for Epidemiologic Studies Depression Scale, commonly designated as the CES-D-10. Within this standardized assessment instrument, a cumulative total score of 10 or higher was established as the threshold indicating the presence of clinically meaningful depressive symptoms.

Incomplete observational records and missing covariate values across the study cohort were imputed using the multiple imputation by chained equations framework, ensuring robust statistical preservation of participant data. Candidate variables were first filtered through least absolute shrinkage and selection operator regression, commonly termed LASSO, and subsequently refined via multivariable logistic regression modeling to isolate the most predictive subset of demographic, functional, and lifestyle characteristics.

The resulting feature selection workflow successfully condensed the patient profile into six essential predictors: self-assessed health status, nightly sleeping duration, the presence of diagnosed memory-related disease, gender, functional capacity as quantified by activities of daily living scores, and chronological age. The researchers also conducted random forest impurity-based and permutation feature-importance analyses.

The researchers evaluated and compared six supervised machine learning algorithms. The evaluated architectures encompassed classical multivariable logistic regression, multilayer perceptron neural networks, K-nearest neighbors, random forest ensembles, XGBoost, and LightGBM. The models were evaluated using 10-fold cross-validation within the training set and an independent test set. To further evaluate stability and minimize sampling bias, the investigators subjected the primary competing algorithms—logistic regression and random forest—to an extensive simulation involving 50 repeated stratified random training-test splits.

The experimental results demonstrated that the conventional logistic regression model achieved the most stable generalization performance across the evaluation benchmarks. In the independent testing cohort, logistic regression attained an area under the receiver operating characteristic curve of 0.713, accompanied by a 95 percent confidence interval spanning from 0.648 to 0.778. Across the 50 repeated stratified random splits, logistic regression consistently outperformed the more complex random forest algorithm, yielding a superior average sensitivity of 0.698 with a standard deviation of 0.060, compared to 0.641 with a standard deviation of 0.036 for random forest. Similarly, logistic regression secured a higher average F1 score of 0.706 with a standard deviation of 0.035, compared to 0.671 with a standard deviation of 0.021 for random forest, while maintaining comparable specificity across both models at 0.606 with a standard deviation of 0.061 versus 0.616 with a standard deviation of 0.046.

Despite the model’s favorable predictive stability and transparent mathematical structure, the authors emphasized key methodological boundaries and clinical caveats regarding its current deployment. The researchers explicitly noted that the logistic regression model must be regarded strictly as a screening-support instrument intended to guide clinical attention, rather than as a definitive diagnostic instrument. Furthermore, the investigative team highlighted an important limitation regarding cohort ascertainment: participant classification was established via self-reported physician diagnoses of pulmonary disease rather than clinical confirmation through spirometry testing. Consequently, the authors concluded that while the model provides a practical framework for community screening, broad clinical adoption will require extensive external validation across independent patient cohorts with spirometry-confirmed chronic obstructive pulmonary disease.

Subject of Research: Machine learning prediction of depressive symptoms in community COPD patients

Article Title: Development and validation of an interpretable risk prediction model for depressive symptoms among community-based patients with chronic obstructive pulmonary disease using machine learning techniques

Article References: Chen, Y., Yu, Y., Yang, D., Chen, X., & Chen, J. (2026). Development and validation of an interpretable risk prediction model for depressive symptoms among community-based patients with chronic obstructive pulmonary disease using machine learning techniques. BMC Pulmonary Medicine. https://doi.org/10.1186/s12890-026-04734-w

Image Credits: AI Generated

DOI: 10.1186/s12890-026-04734-w

Keywords: Chronic obstructive pulmonary disease, Depressive symptoms, China Health and Retirement Longitudinal Study, Logistic regression, Random forest, Machine learning, CES-D-10, Feature selection, LASSO regression, Activities of daily living, Self-assessed health, Sleep duration, Memory disease, Risk prediction model, Screening tools, XGBoost, LightGBM, Multilayer perceptron