Premium
A predictive, modeling‐based screening tool to enrich amyloid beta positivity in a cognitively normal sample
Author(s) -
TrueloveHill Monica,
Erus Guray,
Pomponio Raymond,
Bashyam Vishnu,
Doshi Jimit,
Habes Mohamad,
Ezzati Ali,
Bilgel Murat,
Resnick Susan M.,
Nasrallah Ilya M.,
Wolk David A.,
Davatzikos Christos
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.045242
Subject(s) - cutoff , logistic regression , population , sample size determination , psychology , medicine , neuroimaging , clinical psychology , statistics , psychiatry , mathematics , physics , environmental health , quantum mechanics
Background As β‐amyloid (Aβ) targeted clinical trials have had questionable benefit in treating patients with early symptomatic Alzheimer’s disease, treatment would presumably be most effective if administered during the disease’s preclinical phase. PET scans, while highly sensitive to AD pathology, are less practical for screening a large, cognitively normal population for evidence of preclinical AD. Recruitment methods that produce an Aβ enriched sample before PET data collection would decrease participant burden and potentially lower costs. Here, we develop a simple predictive model using subject demographics, structural MRI, and APOE 4 status. Using the model’s estimates of one’s probability of Aβ positivity, we significantly increase the proportion of Aβ+ subjects in a cognitively normal sample. Method Data from 875 participants from the Baltimore Longitudinal Study of Aging, Penn Memory Center and the Alzheimer’s Disease Neuroimaging Initiative database were separated into a discovery set including AD, MCI, and cognitively normal participants ( n = 673) and a cognitively normal replication set ( n = 202). Aβ prediction performance of a logistic regression model was compared across multiple predictor sets. Regression‐based probability values were used to threshold the data, i.e., participants whose values fell below a cutoff were excluded from the thresholded sample. The proportion of Aβ+ individuals between the restricted and complete samples was compared across multiple cutoffs. Result Applying a probability cutoff of 0.4 yielded samples with significantly higher proportions of Aβ+ subjects in 3 of the 5 tested models (Table 1). The Lesions model achieved the highest sample enrichment; 40.54% of the thresholded sample were Aβ+, significantly higher than the base rate of 28.22%, χ2 (1) = 8.32, p = .004 (Figure 1). However, there were no significant differences between the enriched sample produced by the Lesions model and those produced by the other significant models, indicating sample enrichment is possible even with different predictors. Conclusion Aβ+ probability metrics produced by predictive models provide a novel method of sample enrichment and can be used as an individualized measure of Aβ classification certainty. Fewer PET scans will be necessary to obtain a sufficient Aβ+ sample, reducing burden on participants and researchers and potentially lowering costs.