Exploring multifractal-based features for mild Alzheimer's disease classification
Purpose: Multifractal applications to resting state functional MRI (rs-fMRI) time series for diagnosing Alzheimer's disease (AD) are still limited. We aim to address two issues: (I) if and what multifractal features are sufficiently discriminative to detect AD from the healthy; (II) if AD classification could be further improved by combining multifractal features with traditional features in this field.
Methods: Rs-fMRI data of 25 AD patients and 38 normal controls were analyzed. A set of multifractal features were systematically investigated. Traditional features in monofractal, linear, and network-based categories were also extracted for comparison and combination. Both support vector machines and multiple kernel learning (MKL) were used to perform classification with individual and combined features.
Results: We identified a multifractal feature, Δf, which has the strongest discriminative power among all the features in our study. Moreover, we found that the classification accuracy could be significantly improved from 69% (by Δf only) to up to 76%, when nonsparse MKL is used to combine Δf with the monofractal feature, Hurst. Finally, we showed that incorporating other multifractal features, α(0), Δα and Pc, could also improve traditional-feature-based AD classification.
Conclusion: Our work demonstrated the potential usefulness of multifractal analysis for AD research, especially when combining with the traditional rs-fMRI features. It contributes to distinguishing AD from NC subjects.