Survival analysis, the infinite Gaussian mixture model, FDG-PET and non-imaging data in the prediction of progression from mild cognitive impairment

TL;DR

This paper introduces a method combining survival analysis and the infinite Gaussian mixture model to identify brain regions in PET scans that predict the progression from mild cognitive impairment to Alzheimer's disease, with improved accuracy when combined with non-imaging data.

Contribution

It presents a novel approach integrating IGMM and Cox regression to identify predictive brain regions and enhance MCI to AD conversion prediction using baseline PET scans and non-imaging data.

Findings

PET imaging data outperforms non-imaging data in prediction accuracy

Combining imaging and non-imaging data improves prediction performance

Identified brain regions facilitate understanding of AD progression

Abstract

We present a method to discover interesting brain regions in [18F] fluorodeoxyglucose positron emission tomography (PET) scans, showing also the benefits when PET scans are in combined use with non-imaging variables. The discriminative brain regions facilitate a better understanding of Alzheimer's disease (AD) progression, and they can also be used for predicting conversion from mild cognitive impairment (MCI) to AD. A survival analysis(Cox regression) and infinite Gaussian mixture model (IGMM) are introduced to identify the informative brain regions, which can be further used to make a prediction of conversion (in two years) from MCI to AD using only the baseline PET scan. Further, the predictive accuracy can be enhanced when non-imaging variables are used together with identified informative brain voxels. The results suggest that PET scan imaging data is more predictive than other non-imaging data, revealing even better performance when both imaging and non-imaging data are combined.