Hybrid multi-modality multi-task learning for forecasting progression trajectories in subjective cognitive decline.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Data Fusion

Background:

• Integrating MRI and PET data for disease progression prediction is challenging due to modality differences.
• Small sample sizes and missing data (PET) are common issues in neurodegenerative disease studies.

Purpose of the Study:

• To develop a hybrid multi-modality multi-task learning (HM²L) framework for forecasting Subjective Cognitive Decline (SCD) progression.
• To address challenges of missing PET data and small sample sizes using cross-domain knowledge transfer.

Main Methods:

• Proposed HM²L framework includes missing PET imputation, multi-modality feature extraction with a softmax-triplet constraint, and attention-based fusion.
• Employed a transfer learning strategy from a large dataset (795 subjects) to two small SCD cohorts (136 subjects).
• Multi-task prediction of category labels and clinical scores (MMSE, GDS).

Main Results:

• HM²L significantly outperformed state-of-the-art methods in jointly predicting SCD category labels and clinical scores.
• Lower Mini-Mental State Examination (MMSE) scores were observed in SCD subjects who progressed to mild cognitive impairment.
• A complex relationship was identified between SCD progression and Geriatric Depression Scale (GDS) scores.

Conclusions:

• The HM²L framework offers an effective approach for multi-modality data fusion and knowledge transfer in neurodegenerative disease research.
• Accurate prediction of SCD progression trajectories is achievable, aiding in early diagnosis and intervention planning.
• Findings highlight the utility of MMSE and GDS in tracking cognitive changes and mood in SCD patients over time.