Predicting Amyloid Positivity from Hippocampal and Entorhinal Cortex Volume and APOE Genotype

Abstract

AbstractBackgroundAccumulation of brain amyloid – an early indicator of AD – is assessed using PET or CSF samples, but such procedures are highly invasive. There is interest in using less invasive data, such as brain MRI, to predict β‐amyloid (Aβ) positivity, as a precursor to more invasive screening.Here we assessed different predictive models based on logistic regression, using readily measured indices from T1‐weighted MRI (T1w). We hypothesized that entorhinal cortex measures would outperform hippocampal measures in predicting amyloid positivity, given its role as a cortical area with early amyloid accumulation.MethodWe analyzed T1w from 724 subjects (Table 1) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). One cut‐off, for determining amyloid levels, was used based on PET cortical SUVR uptake (Αβ_1) determined by either mean 18F‐florbetapir (Aβ+ defined as >1.11) or florbetaben (Aβ+ defined as >1.20), normalized by using whole cerebellum reference region. After registering the T1w image to a common template, we split the data into independent training and testing sets. A logistic regression classifier was used, and test performance was assessed using balanced accuracy to compare the different predictive models.ResultFigure 1 shows classification results from the models. The balanced accuracy was highest (74.9%) for the logistic regression model including both left and right entorhinal cortex volumes, APOE status, age, and sex. This model outperformed models using hippocampal volume as the main imaging predictor (balanced accuracy 70.8%).ConclusionOptimal accuracy was obtained when predicting Aβ positivity from brain MRI using entorhinal cortical measures. Future work will test deep learning methods for Aβ prediction based on paired training data, and will use voxel‐based data and features from other imaging modalities, such as diffusion MRI, to predict amyloid status.