Association of Deep Learning-Derived Temporalis Sarcopenia with Mortality in Acute Ischemic Stroke

BackgroundSarcopenia is associated with mortality and morbidity following acute ischemic stroke (AIS), but the diagnosis requires specialized equipment or time-consuming assessments. Computed tomography (CT) measures of temporal muscle volume (TMV) and density (TMD) can be opportunistically measured from existing scans and automated using deep learning (DL). This study sought to demonstrate the incremental prognostic value of DL-derived TMV and TMD from CT scans on mortality and length of stay (LOS) in AIS. MethodsIn this retrospective, single-centre cohort study, consecutive AIS patients admitted from 2014 to 2023 were included. Admission CT scans were retrieved alongside clinical data from electronic health records. TMV and TMD were quantified by a novel DL model and represented as continuous or trichotomous categorical variables. TMV and TMD thresholds were derived in a cohort of 50 healthy adults and used to classify AIS patients as non-sarcopenic, pre-sarcopenic, or sarcopenic. The primary outcome was 30-day all-cause mortality. Secondary outcomes were 365-day all-cause mortality and LOS. Multivariable logistic and linear regression were used. ResultsThe cohort consisted of 2285 patients with 1151 (50%) females, and a mean (SD) age of 74.7 (13.7) years. Based on TMV and TMD, 877 patients (38%) were non-sarcopenic, 838 (37%) pre-sarcopenic, and 570 (25%) sarcopenic. Adjusted ORs for 30-day mortality were 2.70 (1.64 to 4.46) and 2.91 (1.72 to 4.91) for pre-sarcopenia and sarcopenia. Adjusted ORs for 365-day mortality were 2.42 (95% CI 1.74 to 3.36) and 2.96 (95% CI 2.09 to 4.17) and the additional days in hospital were 2.79 (1.69 to 3.98) and 3.26 (2.00 to 4.64) for pre-sarcopenia and sarcopenia. The association between CT-derived sarcopenia and mortality was preserved after adding HFRS to the models. ConclusionsTMV and TMD extracted using a novel DL model were incrementally predictive of AIS mortality. These metrics may be used to refine risk estimates, inform shared decision-making, and individualize treatment plans.