Classification of Alzheimer's Disease, Mild Cognitive Impairment, and Cognitively Unimpaired Individuals Using Multi-feature Kernel Discriminant Dictionary Learning

Qing Li, Xia Wu, Lele Xu, Kewei Chen, Li Yao, Alzheimer's Disease Neuroimaging InitiativeFrontiers Comput. Neurosci 2017 · OpenReview

TL;DR

This paper introduces a novel machine learning approach, Multi-feature Kernel Discriminant Dictionary Learning (MKSCDDL), to accurately classify individuals with Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), and cognitively unimpaired (CU) states using multimodal neuroimaging data. MKSCDDL demonstrates superior classification accuracy and efficiency compared to existing methods, offering a promising tool for early disease diagnosis.

What Problem Does This Paper Address?

Accurately distinguishing between Alzheimer's Disease (AD), its early prodromal stage Mild Cognitive Impairment (MCI), and healthy cognitive aging (Cognitively Unimpaired - CU) is crucial for timely diagnosis and intervention. Traditional diagnostic methods can be subjective and may not detect the disease in its earliest stages. Neuroimaging techniques like structural Magnetic Resonance Imaging (sMRI), FDG-PET, and florbetapir-PET offer objective data, but effectively integrating information from these diverse modalities to achieve high diagnostic accuracy remains a significant challenge. This paper aims to address this by developing a robust classification framework that leverages the complementary information from multiple imaging types.

Methods

The core of this work is the Multi-feature Kernel Discriminant Dictionary Learning (MKSCDDL) algorithm, an extension of their previously developed Supervised within-Class-Similar Discriminative Dictionary Learning (SCDDL). MKSCDDL builds upon SCDDL by incorporating a kernel-based approach to handle multi-modal data and a "within-Class-similarity" objective function. This means it not only learns a dictionary of features that can discriminate between classes but also encourages features within the same class to be similar, thereby improving robustness. The "kernel" aspect allows MKSCDDL to effectively fuse information from different imaging modalities (sMRI, FDG-PET, florbetapir-PET) by mapping them into a higher-dimensional space where their relationships can be better captured. The algorithm is designed to optimize for classification accuracy while also considering the sparsity of representations and the similarity of features within a class.

Datasets

The study utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. This is a well-established, large-scale, longitudinal study that collects multimodal data from individuals with AD, MCI, and CU. The specific dataset included:

  • 113 AD patients
  • 110 MCI patients
  • 117 CU subjects

The data-acquisition interval for the three modalities (sMRI, FDG-PET, florbetapir-PET) for each subject was within four months. Subjects were matched for age, education, and gender.

Key Insights and Ideas

  • Multimodal Fusion is Key: The study strongly suggests that combining information from different neuroimaging modalities (sMRI, FDG-PET, florbetapir-PET) significantly enhances classification performance compared to using any single modality alone. This highlights the complementary nature of these imaging techniques in capturing different aspects of brain pathology.
  • MKSCDDL's Effectiveness: The proposed MKSCDDL algorithm, with its kernel-based multi-feature fusion and within-class similarity objective, proves to be highly effective in distinguishing between AD, MCI, and CU states. Its ability to learn discriminative features and group similar ones is a critical factor.
  • Early Detection Potential: The high accuracy achieved, particularly in distinguishing MCI (an early stage of AD) from CU, underscores the potential of MKSCDDL for early diagnosis, which is crucial for initiating timely interventions and potentially slowing disease progression.
  • Biomarker Discovery: The research also hints at the possibility of identifying key brain regions (biomarkers) that are most discriminative for each classification task, suggesting future avenues for more targeted diagnostic approaches.

Key Results

The MKSCDDL algorithm achieved impressive classification accuracies:

  • AD vs. CU: 98.18% accuracy
  • MCI vs. CU: 78.50% accuracy
  • AD vs. MCI: 74.47% accuracy

These results were consistently superior to several other state-of-the-art classification methods, including MKL, JRC, mSRC, and mSCDDL, across all classification pairs. Furthermore, MKSCDDL demonstrated significantly faster testing times compared to some of these methods, making it more computationally efficient for practical applications.

Limitations and Future Work

  • Generalizability: While the ADNI dataset is comprehensive, the authors acknowledge that further validation on independent datasets would be beneficial to confirm the generalizability of their findings.
  • Biomarker Specificity: While the study identified potential biomarkers, further research is needed to fully understand their specific roles and to refine their use in clinical practice.
  • Parameter Optimization: The performance of MKSCDDL relies on optimal parameter settings. While they performed cross-validation, exploring more advanced optimization techniques could further enhance performance.
  • Longitudinal Prediction: The current study focuses on cross-sectional classification. Future work could explore MKSCDDL's ability to predict disease progression over time using longitudinal data.

Summary updated

Abstract

Accurate classification of either patients with Alzheimer's disease (AD) or patients with mild cognitive impairment (MCI), the prodromal stage of AD, from cognitively unimpaired (CU) individuals is important for clinical diagnosis and adequate intervention. The current study focused on distinguishing AD or MCI from CU based on the multi-feature kernel supervised within-Class-similar discriminative dictionary learning algorithm (MKSCDDL), which we introduced in a previous study, demonstrating that MKSCDDL had superior performance in face recognition. Structural magnetic resonance imaging (sMRI), fluorodeoxyglucose (FDG) positron emission tomography (PET), and florbetapir-PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database were all included for classification of AD vs. CU, MCI vs. CU, as well as AD vs. MCI (113 AD patients, 110 MCI patients, and 117 CU subjects). By adopting MKSCDDL, we achieved a classification accuracy of 98.18% for AD vs. CU, 78.50% for MCI vs. CU, and 74.47% for AD vs. MCI, which in each instance was superior to results obtained using several other state-of-the-art approaches (MKL, JRC, mSRC, and mSCDDL). In addition, testing time results outperformed other high quality methods. Therefore, the results suggested that the MKSCDDL procedure is a promising tool for assisting early diagnosis of diseases using neuroimaging data.
BibTeX
@inproceedings{li2017classification,
  author = {Qing Li and Xia Wu and Lele Xu and Kewei Chen and Li Yao and Alzheimer's Disease Neuroimaging Initiative},
  title = {{Classification of Alzheimer's Disease, Mild Cognitive Impairment, and Cognitively Unimpaired Individuals Using Multi-feature Kernel Discriminant Dictionary Learning}},
  booktitle = {Frontiers Comput. Neurosci},
  year = {2017},
  url = {https://openreview.net/forum?id=ejFwTyFV_vi},
}
Figure