Memorability analysis for diagnostic photographs in cognitive assessment: Linking behavioral performance with biomarker status. (31st December 2021)

• Record Type: Journal Article
• Authors: Bai, Yuetong; Schütze, Hartmut; Berron, David; Ziegler, Gabriel; Spottke, Annika; Buerger, Katharina; Schneider, Anja; Peters, Oliver; Priller, Josef; Wiltfang, Jens; Laske, Christoph; Teipel, Stefan J.; Wagner, Michael; Jessen, Frank; Düzel, Emrah; Bainbridge, Wilma A.
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 6
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-12-31
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.052476
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Recent work has found that memorability, an intrinsic image property predictive of memory, shows consistency across healthy controls (HC), subjective cognitive decline (SCD) and mild cognitive impairment (MCI). Looking at memorability pattern differences between groups, certain images in which HC outperform MCI or SCD individuals can also better predict group membership with a reduced number of images. A key question is whether these diagnostic images reflect a strong link between behavioral performance and biomarker status. Method: Behavioral and biomarker data was collected from 232 participants (64 HC, 99 SCD, 48 MCI, 21 AD) from the DZNE Longitudinal Cognitive Impairment and Dementia (DELCODE) study. Participants were tested for recognition of images from a pool of 835 scene images. Aβ42, total tau, phospho‐tau, Aβ42/Aβ40, and Aβ42/phospho‐tau were collected as biomarkers. We computed linear models predicting corrected recognition with each biomarker as a separate regressor. Using hold‐out cross‐validation (80% training) across 1000 iterations, we also identified three subsets of 50 images in which performance differences between HC and MCI were the highest ("top subset"), lowest ("bottom subset"), or close to zero ("middle subset"). We tested these subsets on the remaining 20% test data, by correlating each biomarker and the average hit rate for that image subset, and comparing their performance with independent t‐tests. Result: All five biomarkers significantly predicted corrected recognition (p<10 ‐5 ) in the linear models (Table 1). For all biomarkers, the correlation coefficients acquired from the top subsets were significantly larger than those acquired from the middle subsets or bottom subsets (Fig. 1). In contrast, the middle and bottom subsets showed small differences that were inconsistent across biomarkers. Conclusion: Our results reveal that images with larger memorability differences between HC and patient groups provide stronger links between behavioral performance and biomarker status. These images are more indicative of neuropathologic changes relevant to AD and are thus more diagnostic. Future study may provide more evidence with functional imaging data.