A machine learning framework with uncertainty‐driven neuroimaging modality selection for data‐efficient prediction of Alzheimer’s Disease

Abstract

AbstractBackgroundEarly prediction of the AD risk for individuals with MCI has important clinical value. Machine Learning (ML) models integrating multi‐modality neuroimaging datasets have shown great promise. Existing approaches either require all the modalities or suffer from performance loss with fewer modalities. We proposed a data‐efficient ML framework, namely the Uncertainty‐driven Modality Selection (UMoS) framework, to sequentially add modalities for each patient on an as‐needed basis while at the same time ensuring high prediction accuracy.MethodThe dataset included 1319 T1‐MRI scans from MCI patients in ADNI and among these 612 additionally had amyloid‐PET. Regional volumetric and thickness measures were computed using FreeSurfer v7.1 from MRI and regional SUVR measures were computed using our in‐house pipeline from amyloid‐PET. MRI is used in the standard of care while amyloid‐PET has less accessibility and is more expensive. Thus, the goal of UMoS was to save a patient from needing PET if doing so did not hurt prediction accuracy. To achieve this, we first trained two elastic net logistic regression models to predict the risk of conversion to AD for MCI patients: one with MRI only; the other one additionally including amyloid‐PET. Uncertainty quantification (UQ) was performed to quantify the predictive uncertainty of the MRI model. Under the UMoS framework, each individual was first predicted using the MRI model. If the predictive uncertainty was high, the individual would be predicted using the MRI+PET model. The uncertainty threshold was selected by balancing the saving of PET in the patient cohort and the prediction accuracy.ResultUMoS achieved 0.851 Area Under the Curve (AUC) with 46.2% patients predicted by the MRI‐only model and 53.8% by the MRI+PET model. If all patients were predicted by the MRI model, the AUC was 0.794. If all patients were predicted by the MRI +PET model, the AUC was 0.875.ConclusionWe proposed an ML framework, UMoS, that used predictive uncertainty of the MRI model to drive the decision that if less‐accessible amyloid‐PET was needed for predicting MCI conversion to AD. This framework allows an individualized approach to make accurate prediction based on the data available and leverage additional data only when needed.