Few-Shot Contrastive Learning for Cross-Task Stroke Prognosis Prediction With Multimodal Data.

Area of Science:

• Medical Imaging
• Machine Learning
• Neurology

Background:

• Stroke outcome prediction is complex due to data heterogeneity and misalignment.
• Current methods lack robustness and generalizability for multimodal stroke data.

Purpose of the Study:

• To develop a robust few-shot learning framework for cross-task stroke prognosis prediction.
• To integrate brain MRI images and structured clinical records for improved outcome prediction.

Main Methods:

• A few-shot contrastive learning framework combining Model-Agnostic Meta-Learning (MAML) and a two-step contrastive strategy.
• Incorporation of self-awareness learning for task-specific features and domain learning for generalization.
• Utilized Misalignment Separation technique for handling inconsistent tabular data and a domain encoder for multimodal inputs.

Main Results:

• The proposed framework demonstrated superior performance compared to SimCLR and conventional supervised methods.
• Successfully integrated cross-task datasets, outperforming existing approaches.
• Showcased enhanced predictive robustness by capturing shared and task-specific knowledge.

Conclusions:

• Multimodal few-shot learning offers a promising approach for robust stroke prognosis prediction, particularly with small sample sizes.
• The framework effectively addresses challenges of data heterogeneity and misalignment in stroke outcome prediction.
• Highlights the potential for improved clinical decision-making in stroke care.