yvsoucom-iterkit: reproducible AutoML hits 0.94 F1 on stroke risk prediction

A new paper from Rui Huang and Lican Huang presents yvsoucom-iterkit, a deterministic and log-driven automated machine learning framework designed for reproducible and interpretable pipeline optimization in healthcare risk prediction. The framework encodes each pipeline as a traceable log entity, enabling detailed analysis of component attribution, interactions, similarity, and cross-seed robustness. The researchers tested yvsoucom-iterkit on the Pima Indians Diabetes and Stroke datasets with over 18,000 pipeline configurations, revealing a structured and partially redundant search space where only a small subset of components governs performance.

Key findings show that augmentation (importance 0.454), model choice (0.198), and imbalance handling (0.101) are the primary drivers on Pima, while imbalance handling dominates Stroke (0.406). Component similarity analysis identified strong redundancy among feature selection variants and augmentation methods. Ensemble models achieved the best results: Weighted-F1 of 0.89 and Macro-F1 of 0.88 on Pima, and Weighted-F1 of 0.94 on Stroke (Macro-F1 lower at 0.67 due to class imbalance). Cross-seed analysis revealed a performance-robustness trade-off, with ensembles showing lower variability (0.023-0.026) than SVM. The study implies effective AutoML can focus on a reduced set of high-impact components, improving efficiency and interpretability.