Abstract
Eye-tracking (ET) metrics obtained during the Visual Short-Term Memory Binding Task (VSTMBT) have shown promise in detecting early and subtle alterations in individuals at risk for, or diagnosed with, Alzheimer's disease (AD) dementia. However, there remains a critical need for robust, automated classification methods capable of delivering affordable digital biomarker solutions for the preclinical detection of AD. This study aimed to address this need. A sample of 100 carriers (89 healthy asymptomatic carriers-HAC and 11 symptomatic familial Alzheimer's disease-FAD) of the E280A mutation in PSEN1 from the widely investigated cohort in Antioquia, Colombia, and 119 healthy controls (Controls HCA: 91 and Controls FAD: 28) participated in the study. The groups were assessed using the novel VSTMBT coupled with ET and an extensive neuropsychological battery. Oculomotor behaviours were recorded using ET, and their analysis was based on Machine Learning classification using Random Forest (RF) Models. Classification accuracy incorporated both true and false positives and negatives. The RF models that incorporated oculomotor behaviours accurately identified FAD (Accuracy = 100%) and HAC (Accuracy = 96%), outperforming classification accuracy based on pure behavioural scores (FAD = 98% and HAC = 73%). The cognitive biomarker drawn from RF models that incorporated oculomotor behaviours accurately detected mutation carriers who inevitably develop FAD and outperformed traditional forms of cognitive assessment. The oculomotor phenotype unveiled here characterizes the preclinical stages of FAD, as it has been identified in most carriers, even those in the still asymptomatic stages.