Machine-learning prediction of post-stroke depression using admission clinical and laboratory features: a derivation study
Background Post-stroke depression (PSD) affects approximately one third of patients after acute ischemic stroke and impairs functional recovery, quality of life and survival, yet no admission-time prediction tool is in routine use. We aimed to derive and internally evaluate machine-learning (ML) models for PSD using routinely available admission variables. Methods We analysed 378 adults with acute ischemic stroke (156 with PSD; 41.3%), randomly split into a training set ( n = 264) and a held-out validation set ( n = 114) in a stratified 70/30 ratio. Logistic regression (LR), random forest (RF) and XGBoost were trained on 21 admission variables. Within the training set, repeated stratified 5-fold cross-validation (10 repeats) provided internal performance estimates; the final models were refit on the full training set and evaluated once on the held-out validation set. Discrimination (area under the receiver operating characteristic curve, AUC), sensitivity, specificity, Brier score and calibration were assessed; SHAP analysis and multivariable logistic regression provided model interpretation. Results XGBoost achieved the highest held-out discrimination (AUC 0.905, 95% CI 0.867–0.933; sensitivity 0.787, specificity 0.821, Brier 0.124), followed by RF (AUC 0.890, 95% CI 0.847–0.922) and LR (AUC 0.849, 95% CI 0.788–0.894); paired DeLong tests showed XGBoost significantly outperformed LR ( P = 0.043) but not RF ( P = 0.40), and RF did not significantly outperform LR ( P = 0.09). Training internal cross-validation AUCs were 0.812 (LR), 0.801 (RF) and 0.792 (XGBoost), a different ordering from the hold-out set. Left-hemisphere stroke, baseline NIHSS, education, homocysteine and sex were the dominant predictors across SHAP and multivariable analyses. XGBoost was reasonably well calibrated on the held-out validation set (calibration slope 1.28, 95% CI 0.96–1.90; intercept − 0.24, 95% CI -0.84 to 0.31; Brier 0.124). Conclusions Admission variables alone support a preliminary, internally evaluated PSD risk-stratification model; among the three algorithms tested, XGBoost gave the highest held-out discrimination, although its advantage over logistic regression was of borderline significance. External validation and probability recalibration remain priorities before clinical deployment.
Authors
- Xiaoyu Ni
- Ling Hu (ORCID: https://orcid.org/0000-0001-7318-5091)
- Qiu Han (ORCID: https://orcid.org/0000-0003-0070-1870)
- Ming Wei (ORCID: https://orcid.org/0000-0003-3392-3939)
- Xiu Yang
Institutions
- Huaian First People’s Hospital (CN)
- Nanjing Medical University (CN)
Publication Details
- Journal
- BMC Medical Informatics and Decision Making
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1186/s12911-026-03961-9
- Primary Topic
- Artificial Intelligence in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00