AutoML in High-Dimensional Biomedical Analysis: A Comparative Study of Predictive Performance and Computational Efficiency Across Genomic and Facial Image Datasets

This study presents the first comprehensive benchmark of five leading AutoML frameworks (TPOT, MLJAR, FLAML, H2O, and AutoGluon) for analyzing ultrahigh-dimensional biomedical datasets. We reveal critical performance-efficiency trade-offs through rigorous testing on two representative tasks, which are cancer classification using RNA-Seq data (more than 20,000 genes) and facial image-based age regression (more than 4,000 features). AutoGluon achieved optimal accuracy (100% cancer classification, MAE=6.37 years for age prediction), while FLAML demonstrated superior computational efficiency (8 times faster than evolutionary approaches). Feature engineering significantly boosted performance: PCA improved classification F1-scores by +0.0123, and CNN embeddings reduced regression errors by 34.5%. Our experiments, spanning 20 distinct framework-task-feature combinations, further highlight that H2O excels in distributed genomic classification, while AutoGluon dominates regression with deep features. Based on these findings, we establish practical selection guidelines balancing data modality, resource constraints, and accuracy requirements. This work bridges a critical gap in AutoML adoption for high-dimensional biomedicine, providing actionable insights for deploying automated pipelines in clinical and research settings.