Meta-Learning-Driven Few-Shot Contrastive Learning for Stroke Prognosis Prediction across Multimodal Datasets

Motivation: The challenge of poor generalization performance with small sample sizes in stroke prognosis prediction tasks, especially due to difficulties in collecting follow-up data. Goal(s): To develop a framework that effectively utilizes small yet related datasets for stroke prognosis prediction, improving generalization and performance on limited data. Approach: The proposed approach is a few-shot contrastive learning framework that integrates a two-step meta-learning training paradigm, capturing domain-specific prior knowledge using both structured and unstructured data. Results: Evaluations on two stroke datasets (341 and 309 patients) show that the proposed framework outperforms SimCLR and traditional supervised methods, demonstrating improved resilience with reduced data availability. Impact: This research advances stroke recovery prediction by enhancing model robustness and generalization with limited data. The framework's ability to integrate diverse datasets could improve clinical decision-making in stroke rehabilitation, addressing a critical gap for accurate predictive modeling in this domain.