Heterogeneous Covariates-Aware Pseudo Supervised Meta-Learning for Few-shot Diabetes Classification

Objective: The limited labeled data hinders the application of medical artificial intelligence technology in the field of diabetes classification. In this paper, a pseudo-label supervised meta-learning algorithm supported by heterogeneous covariates data is proposed to implement diabetes classification tasks with fewer labeled samples. Methods: First, clustering algorithms are employed to generate pseudo labels of samples, which are further used to create multiple pseudo-supervised tasks for meta-learning within the framework of few-shot learning. Second, the time and date features of dynamically monitored glucose data are extracted as dynamic covariates, while the physiological indicators from medical single sampling serve as static covariates. By incorporating these heterogeneous covariates, the model inputs are enriched from multiple perspectives, compensating for the homogeneity deficiency of data and providing complementary information. Finally, a pseudo-supervised meta-learning algorithm is proposed to learn the data features supported by heterogeneous covariates in a task-driven manner. The optimal model is then fine-tuned on downstream real diabetes classification tasks, enabling rapid adaptation to unseen new tasks. Results: The proposed algorithm is thoroughly evaluated using clinical data, achieving an accuracy of 95.994% and an F1 score of 91.261%. Conclusion: The proposed method remains preferable for diabetes classification when compared to the state-of-the-art methods. Significance: The approach offers an effective strategy for diabetes classification tasks with incomplete and limited labeled data.