Predicting the pace of MCI progression to AD based on multi‐modality neuroimaging data by a novel hybrid ordinal learning algorithm

Abstract

AbstractBackgroundMachine learning (ML) has shown great promise for integrating multi‐modality neuroimaging datasets to predict the risk of converting to AD for individuals with Mild Cognitive Impairment (MCI). Most existing work aims to classify MCI patients into converters versus non‐converters. The limitation is a lack of granularity in differentiating MCI patients who convert at different paces. Pace prediction has important clinical values by allowing for more personalized interventional strategies, better preparing patients and their caregivers, and facilitating patient selection in clinical trials. We proposed a novel Hybrid Ordinal Learning (HOL) algorithm for pace prediction of MCI conversion to AD.MethodThis study focused on the amyloid‐positive MCI patient cohort in ADNI for the pace prediction, which included 282 samples. The datasets included baseline T1‐MRI, amyloid‐PET, MMSE, CDR, age, gender, years of education, and APOE status. ROI‐based structural measures were computed from T1‐MRI using FreeSurfer v7.1 and regional SUVR measures were computed using our in‐house pipeline from amyloid‐PET. HOL was trained to predict the conversion to AD for each MCI patient into four ordinal classes representing four fast‐to‐slow paces of conversion, e.g., conversion in the first year, not in the first but in the second year, not in the first two years but by the fifth year, and not in the first five years, respectively. The novelty of HOL compared to existing ordinal classification algorithms was its capability of leveraging imperfectly‐labeled samples in the training set to build a robust learner.ResultUnder cross validation, HOL achieved 0.83 accuracy, balanced across the different paces, based on a combination of 314 features from MRI, PET, and clinical data, while the conventional ordinal SVM classifier achieved only 0.68 accuracy.ConclusionWe demonstrated that a novel ML algorithm, HOC, could achieve a high accuracy in predicting the pace of conversion to AD for MCI patients. Future research is needed to validate the research finding with larger datasets.