DoctorAgents: LLM agents auto-build clinical ML pipelines, beating AutoML

Building reliable machine learning pipelines for clinical data is notoriously difficult—data is scarce, heterogeneous, and temporal. Traditional AutoML systems brute-force search predefined spaces, wasting compute and lacking explicit reasoning or memory. DoctorAgents, a new framework from a team including Yue Li at the University of Waterloo, reframes AutoML from exhaustive search to reasoning-driven refinement. Its core innovation: a trio of specialized large language model agents—one generates pipeline configurations, one validates them against data, and another refines failures. These agents communicate through natural-language feedback, allowing the system to backpropagate textual gradients for targeted updates without exploring the whole search space.

In experiments across diverse clinical tasks, DoctorAgents consistently outperformed established AutoML baselines, all while producing more interpretable, task-specific representations. Crucially, the framework is designed for small temporal datasets—like longitudinal patient records—where standard deep learning and AutoML often overfit. By combining LLM reasoning with explicit memory of past iterations, DoctorAgents adapts pipelines based on what failed and why, rather than randomly retrying. The 34-page paper (arXiv:2608.05375) includes five figures and benchmarks against multiple baselines. This approach could significantly reduce the human effort needed to build robust ML systems for high-stakes medical decisions, while making the pipeline design process more explainable to clinicians and regulators.