Meta-Learning for Automated Algorithm Selection in Healthcare: Enhancing Clinical Outcomes through Data-Driven Recommendation

The integration of machine learning (ML) algorithms in healthcare is transforming diagnostics, treatment planning, and patient management. However, the complexity and diversity of these algorithms, coupled with heterogeneous medical data, challenge practitioners without ML expertise. Traditional trial-and-error algorithm selection is inefficient and often suboptimal. This paper introduces a meta-learning framework for automated algorithm recommendation in medical applications. We extract a rich set of meta-features from 80 healthcare datasets, including statistical, information-theoretic, structural, landmarking, and complexity metrics. These meta-features train a Ridge regression meta-learner to model relationships between meta-features and the performance of eight classification algorithms. For any new healthcare task, the meta-learner predicts the most suitable algorithm based on the dataset’s meta-features. Our framework, evaluated on diverse real-world medical datasets, outperforms three other meta-learning systems in algorithm selection. By automating this process, we aim to enhance access to advanced ML, streamline workflows, and accelerate ML research translation into clinical practice, ultimately improving patient care and outcomes.