Abstract
Multimodal heterogeneous data, such as structural magnetic resonance imaging (MRI), positron emission tomography (PET), and cerebrospinal fluid (CSF), are effective in improving the performance of automated dementia diagnosis by providing complementary information on degenerated brain disorders, such as Alzheimer's prodromal stage, i.e., mild cognitive impairment. Effectively integrating multimodal data has remained a challenging problem, especially when these heterogeneous data are incomplete due to poor data quality and patient dropout. Besides, multimodal data usually contain noise information caused by different scanners or imaging protocols. The existing methods usually fail to well handle these heterogeneous and noisy multimodal data for automated brain dementia diagnosis. To this end, we propose a high-order Laplacian regularized low-rank representation method for dementia diagnosis using block-wise missing multimodal data. The proposed method was evaluated on 805 subjects (with incomplete MRI, PET, and CSF data) from the real Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. Experimental results suggest the effectiveness of our method in three tasks of brain disease classification, compared with the state-of-the-art methods.
Introduction
Diagnosing Alzheimer’s Disease (AD) and its precursor, Mild Cognitive Impairment (MCI), is a complex task often likened to solving a puzzle with missing and noisy pieces. Researchers typically rely on multimodal data—specifically structural MRI, PET scans, and CSF biomarkers—to capture a comprehensive view of brain degeneration. However, real-world clinical data is frequently "block-wise missing" due to high costs, invasive procedures, or patient attrition. Furthermore, data collected from different scanners and protocols introduces significant noise.
Most existing automated diagnosis tools struggle to integrate these heterogeneous sources effectively. They often rely on pairwise relationships between subjects, failing to capture the complex, high-order similarities shared among groups of patients. To address these challenges, this study introduces hLRR (high-order Laplacian regularized Low-Rank Representation), a framework designed to handle incomplete and noisy multimodal data by leveraging hypergraph theory and low-rank representation.
Methods
The hLRR method operates through a three-stage pipeline designed to maximize the utility of available data while filtering out noise.
- Multimodal Data Grouping: Rather than discarding subjects with incomplete records, the framework partitions the cohort into seven distinct groups based on the availability of MRI, PET, and CSF data.
- The hLRR Model: The core of the method uses Low-Rank Representation (LRR) to decompose input data into a clean, low-rank structure and a separate noise component. Simultaneously, a High-Order Laplacian (Hypergraph) is incorporated. Unlike standard graphs that connect pairs of subjects, hypergraphs use "hyperedges" to connect multiple subjects, capturing the intricate "neighborhood" structures of disease manifestation across the groups.
- Ensemble Classification: A Support Vector Machine (SVM) is trained for each of the seven data groups. For a new patient, the model identifies the available modalities and employs a majority-voting strategy across the relevant SVMs to produce a final diagnosis.
Datasets
The study utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort. The dataset included 805 subjects categorized into Normal Controls (NC), MCI, and AD. While MRI data was complete for all subjects, PET and CSF data were missing for approximately 51% and 49% of the cohort, respectively.
Results and Discussion
The hLRR method was evaluated across three classification tasks: AD vs. NC, AD vs. MCI, and MCI vs. NC.
- Superior Accuracy: In the AD vs. NC task, hLRR achieved an accuracy of 94.42%, surpassing deep learning benchmarks like Stacked Autoencoders (91.37%).
- Early Diagnosis: The model showed significant promise in the challenging MCI vs. NC task, reaching 85.32% accuracy, a substantial improvement over traditional methods that typically range between 65% and 75%.
- Robustness and Sensitivity: The model demonstrated high reliability with an AUC of 95.01% for AD vs. NC and a sensitivity of 99.50% in the AD vs. MCI task, indicating it rarely misses AD cases.
The "aha moment" of this research lies in the use of hypergraphs. By recognizing that groups of patients share subtle degenerative patterns that pairwise comparisons miss, the model captures the non-linear progression of dementia more effectively. Additionally, by explicitly calculating an error matrix through LRR, the model becomes robust to variations in imaging quality.
Conclusion
The hLRR framework offers a robust solution for dementia diagnosis using messy, real-world multimodal data. By grouping subjects based on data availability rather than using potentially biased imputation methods, and by employing high-order relationships, it sets a new standard for accuracy in early-stage detection. Future work will focus on optimizing the computational cost of hypergraph construction and extending the model to other neurodegenerative disorders such as Parkinson’s disease.
Summary updated