SysML: adaptive recommendation system for heterogeneous biomedical data preprocessing and modeling workflows.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Machine Learning

Background:

• High-dimensional omics data in biomedicine requires robust computational frameworks.
• Current trial-and-error approaches for selecting analytical workflows reduce efficiency and reproducibility.
• There is a need for systematic benchmarking of algorithms and preprocessing techniques.

Purpose of the Study:

• To systematically benchmark algorithm-preprocessing combinations for common biomedical data challenges.
• To identify optimal machine learning workflows for small sample sizes, missing values, and class imbalance.
• To develop a data-adaptive workflow recommendation platform for biomedical research.

Main Methods:

• Benchmarking hundreds of algorithm-preprocessing combinations.
• Evaluating performance across datasets with small sample sizes, missing values, and class imbalance.
• Developing and validating the SysML web-based platform for workflow recommendations.

Main Results:

• Tree-based models (Gradient Boosting Decision Tree, XGBoost, Random Forest) excel with small samples and missing data.
• Partial Least Squares Discriminant Analysis (PLS-DA) is effective for imbalanced classes.
• K-means and DBSCAN are robust to moderate missingness (<10%), but performance degrades with higher missingness.
• SysML demonstrated improved model performance and workflow efficiency on real-world biomedical datasets.

Conclusions:

• Adaptive data preprocessing is critical for reliable and reproducible machine learning in biomedicine.
• Algorithm choice alone is insufficient; data characteristics dictate optimal workflow selection.
• The SysML platform supports data-driven decision-making for biomedical data analysis.