Comparison of the additive value of blood-based and retinal biomarkers to clinical features in predicting cognitive decline

Abstract

Background: Retinal imaging features, and blood-based biomarkers may provide insights into cognitive neuropathology. However, the combined utility of retinal imaging features, and blood-based biomarkers for predicting cognitive decline remains under-investigated. Methods: A cohort of memory-clinic subjects (age 71.86 ± 7.91 years, 48.4% male) were followed for 5 years, with annual Clinical Dementia Rating Scale Sum-of-Boxes(CDR-SB) assessments. Cognitive decline was defined as a CDR-SB ≥ 2 increment from baseline. Elastic net regression of 71 multimodal variables for predicting cognitive decline was performed comprising:8 clinical factors, 50 retinal fundoscopic features, and 13 blood-based biomarkers (Ergothioneine, pTau-181, NfL, insulin, leptin, IL-6, IL-8, TNF, NGF, thrombomodulin, GDF-15, hs-cTnT and NT-proBNP). This yielded a set of 13 variables(4 clinical factors, 5 blood-based biomarkers, and 4 retinal fundoscopic indices) which was used for model construction. Predictive models for cognitive decline were built using:1)Clinical variables, 2)Clinical+Retinal variables 3)Clinical+Blood variables, and 4)Clinical+Retinal+Blood variables. Predictive performance for cognitive decline was evaluated using cross-validated AUROC(area under receiver operating characteristic curve). Results: Here we show, of 306 subjects followed-up for 60.0 months [Interquartile range 48.0–60.0], 104 (33.9%) display cognitive decline. The AUROC of the clinical model for cognitive decline is 0.691[0.629-0.754]. The addition of retinal features to the clinical model does not provide additive risk stratification(AUROC 0.680, [0.616-0.743],p = 0.486), but the addition of blood-based biomarkers improves the discriminative ability of the clinical model for cognitive decline(AUROC 0.748,[0.691-0.754],p = 0.016). A combination of clinical+retinal+blood variables yields the highest predictive ability for cognitive decline(AUROC 0.752 [0.695-0.809]). Conclusions: A multimodal model comprising blood, retinal and clinical variables may represent an accessible, non-invasive tool for predicting cognitive decline.