Abstract
ABSTRACT
Background
Stroke imposes substantial mortality and socioeconomic burden in India, particularly among younger adults. However, prospective evidence on the prognostic value of baseline clinical and radiomics features for mortality prediction remains limited.
Objective
To assess mortality at 28, 90, 180, and 365 days after stroke; characterize survival patterns across stroke subtypes and vascular territories; and identify baseline demographic, clinical, and radiomics predictors of mortality.
Methods
This prospective cohort study recruited consecutive patients with stroke admitted to a tertiary care hospital in Mangalore, India. Demographic, clinical, laboratory, and radiomics features were collected at admission, and mortality was assessed at 28, 90, 180, and 365 days. Kaplan–Meier analysis was used to compare survival across stroke subtypes and vascular territories, while Cox proportional hazards regression was used to identify independent baseline predictors of mortality.
Results
Among 63 confirmed stroke cases, 38 (60%) had ischemic stroke (IS) and 25 (40%) had haemorrhagic stroke (HS). Within the middle cerebral artery (MCA) territory, HS patients demonstrated better survival than IS patients due to a greater burden of adverse baseline characteristics such as previous stroke, atrial fibrillation, and higher lesion volume in the IS-MCA subgroup. The combined clinical-radiomics model demonstrated superior prognostic performance than individual models, with higher concordance indices and consistently lower Akaike Information Criterion values. In the combined model, stroke severity (aHR = 2.35, 3.0, 2.93, 2.41) and DWI texture heterogeneity of lesion (aHR = 2.35, 2.11, 2.24, 2.10) were the strongest predictors of mortality across all time points. The CT attenuation heterogeneity (aHR = 1.81, 28 days; 1.76, 90 days; 1.80, 180 days) of lesion and age (aHR = 1.48, 365 days) were also significant predictors.
Conclusion
The radiomics features provide critical complementary information to clinical parameters for predicting 28-day mortality, unlike during mid- to long-term follow-up where their incremental contribution appears to be limited. Our results may have important implications in resource-constrained settings for supporting personalized post-stroke management and rehabilitation.