Early post-stroke depression is clinically heterogeneous, and an admission-time estimate should respect both lesion anatomy and the ordered dependence of depressive symptoms. Whether an explicit lesion-to-central-symptom-to-peripheral-symptom cascade adds predictive information beyond direct lesion-aware regression is unknown. A digital cohort of 17,220 profiles was generated from prespecified demographic, vascular, neurological, anatomical, and 13-symptom constraints. Six deep-structure indicators represented the right and left basal ganglia, right and left internal capsules, right external capsule, and right insula. A lesion–symptom probability cascade first estimated depressed mood, psychiatric anxiety, and loss of interest; estimated the remaining ten symptoms from admission variables and out-of-fold central-symptom probabilities; and then estimated the probability of a Hamilton Depression Rating Scale total above 7. Clinical logistic regression, lesion-augmented logistic regression, extra trees, and histogram boosting were evaluated with the same repeated outer validation splits. Discrimination, average precision, Brier loss, calibration intercept and slope, operating characteristics, prevalence shifts, probability attenuation, and subgroup consistency were examined. Lesion-augmented logistic regression attained an area under the receiver-operating curve (AUROC) of 0.6912 and average precision of 0.7317. The cascade attained an AUROC of 0.6911, average precision of 0.7316, Brier loss of 0.2196, calibration intercept of \(-0.001\), and calibration slope of 0.980. Its AUROC therefore differed from lesion-augmented logistic regression by \(-0.0001\). Right basal-ganglia involvement produced the largest absolute risk difference (36.4 percentage points, bootstrap 95% interval 34.7–38.4). Cascade discrimination remained between 0.677 and 0.697 across prespecified subgroups and was stable when outcome prevalence ranged from 0.35 to 0.65, whereas probability attenuation predictably distorted calibration slopes. Under aggregate-calibrated conditions, modeling a symptom cascade did not improve discrimination beyond direct lesion-aware regression. Deep-structure anatomy accounted for the measurable gain over clinical variables, while the cascade supplied an interpretable decomposition and well-calibrated probabilities. Prospective external validation with individual imaging and symptom trajectories is required before clinical use.
Early post-stroke depression is clinically heterogeneous, and an admission-time estimate should respect both lesion anatomy and the ordered dependence of depressive symptoms. Whether an explicit lesion-to-central-symptom-to-peripheral-symptom cascade adds predictive information beyond direct lesion-aware regression is unknown. A digital cohort of 17,220 profiles was generated from prespecified demographic, vascular, neurological, anatomical, and 13-symptom constraints. Six deep-structure indicators represented the right and left basal ganglia, right and left internal capsules, right external capsule, and right insula. A lesion–symptom probability cascade first estimated depressed mood, psychiatric anxiety, and loss of interest; estimated the remaining ten symptoms from admission variables and out-of-fold central-symptom probabilities; and then estimated the probability of a Hamilton Depression Rating Scale total above 7. Clinical logistic regression, lesion-augmented logistic regression, extra trees, and histogram boosting were evaluated with the same repeated outer validation splits. Discrimination, average precision, Brier loss, calibration intercept and slope, operating characteristics, prevalence shifts, probability attenuation, and subgroup consistency were examined. Lesion-augmented logistic regression attained an area under the receiver-operating curve (AUROC) of 0.6912 and average precision of 0.7317. The cascade attained an AUROC of 0.6911, average precision of 0.7316, Brier loss of 0.2196, calibration intercept of \(-0.001\), and calibration slope of 0.980. Its AUROC therefore differed from lesion-augmented logistic regression by \(-0.0001\). Right basal-ganglia involvement produced the largest absolute risk difference (36.4 percentage points, bootstrap 95% interval 34.7–38.4). Cascade discrimination remained between 0.677 and 0.697 across prespecified subgroups and was stable when outcome prevalence ranged from 0.35 to 0.65, whereas probability attenuation predictably distorted calibration slopes. Under aggregate-calibrated conditions, modeling a symptom cascade did not improve discrimination beyond direct lesion-aware regression. Deep-structure anatomy accounted for the measurable gain over clinical variables, while the cascade supplied an interpretable decomposition and well-calibrated probabilities. Prospective external validation with individual imaging and symptom trajectories is required before clinical use.
