Classification of F-Flutemetamol scans in cognitively normal older adults using machine learning trained with neuropathology as ground truth.

PURPOSE

End-of-life studies have validated the binary visual reads of F-labeled amyloid PET tracers as an accurate tool for the presence or absence of increased neuritic amyloid plaque density. In this study, the performance of a support vector machine (SVM)-based classifier will be tested against pathological ground truths and its performance determined in cognitively healthy older adults.

METHODS

We applied SVM with a linear kernel to an F-Flutemetamol end-of-life dataset to determine the regions with the highest feature weights in a data-driven manner and to compare between two different pathological ground truths: based on neuritic amyloid plaque density or on amyloid phases, respectively. We also trained and tested classifiers based on the 10% voxels with the highest amplitudes of feature weights for each of the two neuropathological ground truths. Next, we tested the classifiers' diagnostic performance in the asymptomatic Alzheimer's disease (AD) phase, a phase of interest for future drug development, in an independent dataset of cognitively intact older adults, the Flemish Prevent AD Cohort-KU Leuven (F-PACK). A regression analysis was conducted between the Centiloid (CL) value in a composite volume of interest (VOI), as index for amyloid load, and the distance to the hyperplane for each of the two classifiers, based on the two pathological ground truths. A receiver operating characteristic analysis was also performed to determine the CL threshold that optimally discriminates between neuritic amyloid plaque positivity versus negativity, or amyloid phase positivity versus negativity, within F-PACK.

RESULTS

The classifiers yielded adequate specificity and sensitivity within the end-of-life dataset (neuritic amyloid plaque density classifier: specificity of 90.2% and sensitivity of 83.7%; amyloid phase classifier: specificity of 98.4% and sensitivity of 84.0%). The regions with the highest feature weights corresponded to precuneus, caudate, anteromedial prefrontal, and also posterior inferior temporal and inferior parietal cortex. In the cognitively normal cohort, the correlation coefficient between CL and distance to the hyperplane was -0.66 for the classifier trained with neuritic amyloid plaque density, and -0.88 for the classifier trained with amyloid phases. This difference was significant. The optimal CL cut-off for discriminating positive versus negative scans was CL = 48-51 for the different classifiers (area under the curve (AUC) = 99.9%), except for the classifier trained with amyloid phases and based on the 10% voxels with highest feature weights. There the cut-off was CL = 26 (AUC = 99.5%), which closely matched the CL threshold for discriminating phases 0-2 from 3-5 based on the end-of-life dataset and the neuropathological ground truth.

DISCUSSION

Among a set of neuropathologically validated classifiers trained with end-of-life cases, transfer to a cognitively normal population works best for a classifier trained with amyloid phases and using only voxels with the highest amplitudes of feature weights.