18F-FDG-PET-based deep learning for predicting cognitive decline in non-demented elderly across the Alzheimer’s disease clinical spectrum

This is the authors' abstract. We don't add key points for this paper.

Background With disease-modifying treatments for Alzheimer's disease (AD), prognostic tools for the pre-dementia stage are needed. This study aimed to evaluate the prognostic value of an 18F-fluorodeoxyglucose-positron emission tomography (18F-FDG-PET)-based deep-learning (DL) model in the pre-dementia stage of mild cognitive impairment (MCI) and normal cognition (NC). Materials and Methods A 18F-FDG-PET-based DL model was developed to classify diagnosis of AD-dementia vs NC using AD Neuroimaging Initiative (ADNI) and Japanese-ADNI (J-ADNI) datasets (n = 756), which provided the degree of similarity to AD-dementia. The prognostic value of the DL output for cognitive decline was assessed in the ADNI MCI (n = 663), J-ADNI MCI (n = 129), and Harvard Aging Brain Study (HABS) NC (n = 274) participants using Cox regression and calculating the integrated area under the time-dependent ROC curves (iAUC), along with clinical information and 18F-FDG-PET standardized uptake value ratio (SUVR). Subgroup analysis in the amyloid-positive ADNI MCI participants was performed using Cox regression and calculating the area under the time-dependent ROC (tdAUC) curves at 4-year follow-up to assess prognostic value of DL output over clinical information, 18F-FDG-PET SUVR, and amyloid PET Centiloids. Results DL output remained independently prognostic among other factors in all three datasets (P < .05 for all by Cox regression). By adding DL output to other prognostic factors, prediction significantly improved in ADNI-MCI (iAUC differences 0.020 [0.007-0.034] before and after adding DL output) and improved without statistical significance in J-ADNI (0.020 [−0.005 to 0.044], and HABS-NC sets (0.059 [−0.003 to 0.126]). DL output showed independent (P = .002 by Cox regression) and significant added prognostic value (tdROC difference 0.019 [<0.001-0.036]) over clinical information, 18F-FDG-PET SUVR, and Centiloids in the amyloid-positive ADNI MCI participants. Conclusion The 18F-FDG-PET-based DL model demonstrated the potential to improve cognitive decline prediction beyond clinical information, and conventional measures from 18F-FDG-PET and amyloid PET and may prove useful for clinical trial recruitment and individualized management.