Network biomarkers for early detection of cognitive impairment using digital clock drawings - Project summary/Abstract Our research aims to enhance the early detection of Alzheimer’s disease and related dementias (ADRD) by developing advanced AI tools using digital neuropsychological (NP) tests. While neuroimaging and blood-based biomarkers are precise, they are often expensive and inaccessible for routine screening. Digital NP tests are easy to administer and offer high-frequency time-stamped pen movement data that traditional methods miss, enabling a nuanced understanding of cognitive function. Despite these advancements, there remains a significant gap in modeling approaches that can leverage such data to predict cognitive impairment accurately. Our focus is on the digital clock drawing test (dCDT), a widely used screening tool in clinical settings that evaluates executive function, visuospatial abilities, and semantic memory. It has two parts: a command test (participants receive verbal/on-screen instructions) and a copy test (participants replicate a pre-drawn clock). Traditional approaches to analyzing the CDT and dCDT rely on extracting handcrafted features, which are subsequently analyzed using statistical or conventional machine learning models. However, these methods fail to fully utilize the rich temporal and spatial information captured by such granular data, often overlooking critical structural and sequential patterns associated with neurodegenerative disorders. The overall objective of our research is to (i) develop novel graph-based representations of dCDT drawings and (ii) build advanced machine learning models to detect cognitive impairment early and accurately. First, we will construct novel and comprehensive network representations of dCDT drawings, capturing the spatial positioning of clock components, their morphological attributes, and relative placement. Using these networks, we will develop a heterogeneous graph attention network with dual attention (GADA) to classify individuals as cognitively intact or impaired. We will then employ graph contrastive learning to enhance the generalizability of our models and identify key nodes and edges contributing to classification decisions. Second, we will develop a signed graph representation to quantify differences in how participants draw command and copy clocks in dCDTs. Using these signed networks, we will build a novel hypothesis testing framework to assess structural differences between test types across cognitive subtypes and develop a novel signed GADA to predict cognitive impairment. This approach leverages within-subject drawing differences for early detection. Our work will utilize data from the Framingham Heart Study, which includes extensive biological, genetic, and phenotypic data, along with digital NP data since 2011. By capturing nuanced aspects of dCDT drawings through advanced and comprehensive graph representations, we can identify subtle indicators of cognitive impairment. This will significantly enhance our ability for early detection of ADRD, providing critical insights for timely clinical interventions. Furthermore, early detection tools based on digital tests enhance accessibility in low-resource settings, particularly for elderly populations and in-home assessments.
