Machine learning-based prediction of amyloid positivity using early-phase F-18 flutemetamol PET

Abstract

Background

Previous studies have suggested that early-phase imaging of amyloid positron emission tomography (PET) may offer information for predicting amyloid positivity.

Objective

This study aimed to evaluate whether early-phase fluorine-18 flutemetamol (eFMM) PET images provide valuable information for predicting amyloid positivity using machine learning (ML) models and whether incorporating clinical and neuropsychological features improves predictive performance.

Methods

In total, 454 patients with mild cognitive impairment (MCI) and Alzheimer's disease (AD) were enrolled and randomly divided into training (n = 354) and test (n = 100) groups. We developed ML models using logistic regression (LR) and linear discriminant analyses (LDA) for predicting amyloid positivity: eFMM features alone (eFMM model), eFMM features combined with clinical features (eFMM + C model), eFMM features combined with neuropsychological features (eFMM + N model), eFMM features combined with both clinical and neuropsychological features (eFMM + C + N model), clinical and neuropsychological features combined (C + N model), and dFMM features alone (dFMM model).

Results

In the test group, the eFMM models achieved areas under the receiver operating characteristic curves (AUROCs) of 0.791 (LR) and 0.779 (LDA). The eFMM + C + N models significantly improved predictive performance, with AUROCs of 0.902 for both LR and LDA, outperforming the eFMM models.

Conclusions

ML predictive models using eFMM PET data demonstrated fair performance in predicting amyloid positivity in patients with MCI and AD. The addition of relevant clinical and neuropsychological features further enhanced the predictive performance of the eFMM models, achieving excellent performance.