A newly developed machine-learning framework can identify acute ischemic stroke patients at elevated risk of developing post-stroke depression by relying exclusively on routine admission data, according to a derivation study published in BMC Medical Informatics and Decision Making. Post-stroke depression impairs functional recovery, long-term survival, and overall quality of life, affecting approximately one-third of individuals following an ischemic event. Despite the substantial clinical burden, standardized predictive instruments for early depression risk remain absent from routine hospital admission workflows.
To address this diagnostic gap, investigators derived and internally evaluated computational risk models using admission parameters that are routinely gathered in standard clinical practice. The derivation investigation analyzed records from 378 adult patients admitted with acute ischemic stroke. Among the full cohort, 156 individuals, representing 41.3 percent of the study population, received a diagnosis of post-stroke depression during clinical follow-up.
The research team partitioned the 378 patients into two distinct groups using a stratified 70 to 30 random split. This procedure allocated 264 patients to an initial training cohort and preserved a held-out validation cohort of 114 patients for testing. Analysts trained three separate computational algorithms, including logistic regression, random forest, and XGBoost, utilizing 21 admission variables that encompassed baseline clinical and laboratory features.
Within the training cohort, investigators performed internal cross-validation using a repeated stratified five-fold design executed over 10 repeats to estimate preliminary performance. Within this internal training phase, logistic regression attained an internal cross-validation area under the receiver operating characteristic curve of 0.812, followed by random forest at 0.801 and XGBoost at 0.792. Following internal cross-validation, the final models were refitted on the full training set and evaluated once on the held-out validation set.
On the held-out validation dataset, the algorithm performance hierarchy shifted from the cross-validation ranking. XGBoost achieved the highest discrimination with an area under the curve of 0.905, bounded by a 95 percent confidence interval of 0.867 to 0.933. XGBoost also recorded a sensitivity of 0.787, a specificity of 0.821, and a Brier score of 0.124. In comparison, random forest attained an area under the curve of 0.890 with a 95 percent confidence interval of 0.847 to 0.922, while logistic regression achieved an area under the curve of 0.849 with a 95 percent confidence interval of 0.788 to 0.894.
Statistical comparisons using paired DeLong tests showed that XGBoost significantly outperformed logistic regression, yielding a P value of 0.043. However, XGBoost did not demonstrate a statistically significant performance advantage over random forest, which produced a P value of 0.40. Furthermore, random forest did not significantly outperform logistic regression, yielding a P value of 0.09. Evaluating model calibration on the held-out validation cohort revealed that XGBoost was reasonably well calibrated, showing a calibration slope of 1.28 with a 95 percent confidence interval of 0.96 to 1.90, and an intercept of minus 0.24 with a 95 percent confidence interval of minus 0.84 to 0.31.
Feature importance assessments conducted through SHAP analysis and multivariable logistic regression identified left-hemisphere stroke, baseline National Institutes of Health Stroke Scale score, education level, homocysteine concentration, and patient sex as the dominant predictors across the derived models. The investigators concluded that routinely available admission variables can successfully support a preliminary risk-stratification system for post-stroke depression. Nonetheless, they emphasized that prospective external validation across diverse patient populations and further probability recalibration remain critical research priorities before these machine-learning tools can be integrated into routine clinical deployment.
Subject of Research: Machine-learning prediction of post-stroke depression risk from routine admission variables
Article Title: Machine-learning prediction of post-stroke depression using admission clinical and laboratory features: a derivation study
Article References: Hu, L., Han, Q., Wei, M., Yang, X., & Ni, X. (2026). Machine-learning prediction of post-stroke depression using admission clinical and laboratory features: a derivation study. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03961-9
Image Credits: AI Generated
DOI: 10.1186/s12911-026-03961-9
Keywords: post-stroke depression, acute ischemic stroke, machine learning, XGBoost, random forest, logistic regression, clinical prediction models, risk stratification, admission laboratory features, stroke recovery, SHAP analysis, NIHSS score, homocysteine, left-hemisphere stroke, discrimination AUC, model calibration, internal validation, medical informatics
Cite Scienmag News
Cassandra Pierce. (November 19, 2026). Machine Learning Predicts Post-Stroke Depression Risk Using Routine Hospital Admission Data. Scienmag. https://scienmag.com/machine-learning-predicts-post-stroke-depression-risk-using-routine-hospital-admission-data/
Cassandra Pierce. "Machine Learning Predicts Post-Stroke Depression Risk Using Routine Hospital Admission Data." Scienmag, 19 November 2026, https://scienmag.com/machine-learning-predicts-post-stroke-depression-risk-using-routine-hospital-admission-data/. Accessed 19 November 2026.
Cassandra Pierce. "Machine Learning Predicts Post-Stroke Depression Risk Using Routine Hospital Admission Data." Scienmag. November 19, 2026. https://scienmag.com/machine-learning-predicts-post-stroke-depression-risk-using-routine-hospital-admission-data/








