A high‐dimensional incomplete‐modality transfer learning method for early prediction of Alzheimer’s disease

Abstract

AbstractBackgroundPrediction of Alzheimer’s disease (AD) risk for individuals with mild cognitive impairment (MCI) provides an opportunity for early intervention. Neuroimaging of different types/modalities has shown promise, but not every patient has all the modalities due to the cost and accessibility constraints. To integrate incomplete multi‐modality datasets, we previously developed a machine learning (ML) model called incomplete‐modality transfer learning (IMTL). We extended the capacity of IMTL to handle high‐dimensional feature sets, namely, HD‐IMTL, to further improve accuracy and robustness.MethodOur dataset included 1319 T1‐MRI scans from MCI patients in ADNI; among them, 1002 had FDG‐PET and 612 had amyloid‐PET. 156 regional volumetric and thickness features were computed from MRI and 83 and 83 regional SUVR features from FDG‐PET and amyloid‐PET, respectively. The dataset is randomly split into training and test sets. The goal of HD‐IMTL was to jointly train 4 ML models to predict MCI conversion to AD in 36 months, with each model based on a certain combination of available modalities, namely, MRI, MRI+FDG, MRI+amyloid, and MRI+FDG+amyloid. These correspond to patient sub‐cohorts that differ in their access to imaging modalities. To handle high‐dimensional features, we employed feature screening to remove uninformative features, performed modality‐wise partial least squares (PLS) to condense remaining features into PLS components, and used correlation tests to select components. To jointly train the 4 ML prediction models, IMTL was used, which is a generative model that uses expectation‐maximization (EM) in joint parameter estimation to facilitate transfer learning. To account for sample imbalance in training, the Synthetic Minority Over‐sampling Technique (SMOTE) was used. The trained models were applied to the test set. 20 training/test splits were repeated and AUCs on the test set were averaged. For comparison, three existing ML models for incomplete‐modality fusion were applied to the same dataset.ResultThe AUCs by HD‐IMTL were 0.802, 0.840, 0.868, and 0.880 for sub‐cohorts with MRI, MRI+FDG, MRI+amyloid, and MRI+FDG+amyloid, respectively. The AUCs by existing methods were lower, with ranges of 0.749‐0.793, 0.769‐0.826, 0.816‐0.863, and 0.832‐0.868.ConclusionHD‐IMTL demonstrated high accuracy in predicting MCI conversion to AD for patients with varying access/availability of imaging modalities.