Signal Architecture of Biomarkers in Machine Learning: A Study on Redundancy and Minimal Efficient Combinations After Myocardial Infarction

This study focuses on machine learning prediction models for biomarkers after myocardial infarction, exploring signal concentration, redundancy, and conditional complementarity, and ultimately constructing a minimal efficient biomarker panel. The study proposes the concept of "signal architecture", focusing on the internal signal distribution patterns of models, aiming to provide more transparent and interpretable machine learning tools for clinical decision-making.

In the diagnosis of cardiovascular diseases, distinguishing between STEMI and NSTEMI is crucial for treatment plans. Traditional methods rely on electrocardiograms and myocardial injury markers, but most machine learning studies only focus on performance metrics (such as AUC) and ignore the internal signal structure of models. This study takes a unique approach by analyzing the distribution of predictive signals at the biomarker level and proposing the core concept of "signal architecture" to enhance the transparency and interpretability of clinical tools.

The study focuses on four key dimensions:

1. Signal concentration: Quantify the change in model performance (ΔAUC) when a biomarker is removed via leave-one-out analysis to identify core driving factors;
2. Redundancy assessment: Identify redundant information using Spearman correlation analysis and feature group ablation experiments;
3. Conditional complementarity: Discover biomarkers with synergistic effects through pairwise combination analysis;
4. Construction of minimal efficient combinations: Find the balance between performance and simplicity through model collapse analysis.

The study is based on a dataset of 152 patients, including 10 core biomarkers covering pathways such as inflammation, matrix remodeling, lipid metabolism, and mineral metabolism:

• MMP-2, MMP-9 (matrix metalloproteinases)
• EMMPRIN (extracellular matrix metalloproteinase inducer)
• IL-6, TNF-α (inflammatory markers)
• FGF-23 (mineral metabolism regulator)
• Klotho (anti-aging protein)
• Total cholesterol, HDL, non-HDL (lipid metabolism indicators)

Methodological innovations include:

• Nested cross-validation: Outer layer for performance evaluation, inner layer for hyperparameter tuning to avoid data leakage;
• Multi-model benchmarking: Comparing random forest, logistic regression, linear SVM, and histogram gradient boosting;
• Permutation importance: Evaluating biomarker contributions, which is more reliable than Gini impurity;
• Stability ranking: Classifying biomarkers by integrating cross-fold stability and single-feature importance;
• Open-source contribution: 14 Jupyter Notebooks covering the entire workflow (from data preprocessing to analysis) to support reproduction and extension.

Key findings:

1. A small number of biomarkers are core driving factors (high ΔAUC, low correlation);
2. Identified redundant biomarkers (e.g., overlapping information between IL-6 and TNF-α);
3. Constructed minimal efficient combinations (maintaining accuracy while reducing testing costs);
4. Model simplification and regularization enhance clinical practicality.

Clinical significance and applications:

• Rapid emergency diagnosis: Distinguish between STEMI/NSTEMI to guide emergency interventional treatment;
• Resource optimization: Reduce redundant tests and lower costs;
• Improved interpretability: Enable doctors to understand biomarker-driven predictions and enhance trust;
• Personalized medicine: Extend to subgroup analysis to support personalized diagnosis.

Limitations: Small sample size (152 cases), lack of external multi-center validation, and based on cross-sectional data. Future directions: Large-scale validation cohorts, external validation, longitudinal tracking dynamic models, and multi-omics integration. Conclusion: This study promotes the application of machine learning in cardiovascular diagnosis. The concept of "less is more" is of great significance for optimizing medical resources, and we look forward to its translation into clinical tools to benefit patients.