A gatekeeper for amyloid status based on FDG‐PET and genetic risk in patients with mild cognitive impairment. (1st February 2022)

• Record Type: Journal Article
• Authors: Doering, Elena; Hönig, Merle C; Bischof, Gerard N; van Eimeren, Thilo; Drzezga, Alexander
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 4
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-02-01
• 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.057433
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Amyloid‐β (Aβ) accumulation is a characteristic hallmark for Alzheimer's disease (AD), which is known to be modified by the APOE genotype. Current diagnostic recommendations for AD include the assessment of amyloid status (A‐status: Aβ+ or Aβ‐) through positron‐emission‐tomography (PET) or invasive lumbar puncture. Amyloid‐PET imaging is costly and currently not generally covered by insurance companies. FDG‐PET, in contrast, is comparably inexpensive, commonly available and often employed to assess extent and localization of neurodegeneration (N‐status) in patients with (mild) cognitive impairment (MCI). We aimed to explore, whether advanced machine learning allows for an approximation of the A‐status from FDG‐PET data and APOE genotype, which could make these biomarkers a gatekeeper in deciding on the need for additional amyloid testing. Method: Models for the approximation of the A‐status were trained for APOE4 carriers (APOE‐c) and non‐carriers (APOE‐nc) in parallel. We included 574 FDG‐PET scans of patients with MCI from the ADNI cohort (185 Aβ‐APOE4‐nc; 107 Aβ+APOE4‐nc; 46 Aβ‐APOE4‐c; 236 Aβ+APOE4‐c). A‐status was determined from established cut‐off values for cerebrospinal fluid and/or amyloid‐PET. Using SPM12, regional FDG‐PET standardized uptake value ratios were extracted from spatially normalized scans for 90 brain regions. Different machine learning classifiers were trained on 70% of the data, while the remaining 30% ( n =85 (APOE4‐c); n =88 (APOE4‐nc)) were reserved for testing. APOE‐c and ApoE‐nc models were trained to maximize precision of Aβ‐ and Aβ+, respectively ("relevant instances"), as well as the percentage of relevant instances identified by a classifier ("recall") by means of the F1/10 ‐score. Finally, we estimated the contribution of individual brain regions towards classification performance with permutation importance. Result: In the test set, amyloid status was predicted with a precision of 85% (APOE4‐c) and 87% (APOE4‐nc) and a recall of 93% (APOE4‐c) and 58% (APOE4‐nc). Patterns of neurodegeneration in temporal and frontal (APOE‐nc) and temporo‐parietal regions (APOE‐c) most heavily influenced the classification process (see Figure 1). Conclusion: We propose a highly precise and biologically plausible gatekeeper for amyloid status in MCI patients with varying genetic risk using FDG‐PET. Further validated, our models could serve to reduce the number of patients requiring amyloid testing to determine AD‐related etiology of ongoing neurodegeneration.