Machine Learning for Predicting Stroke Risk Stratification Using Multiomics Data: Systematic Review

Area of Science:

• Integrative bioinformatics and computational biology
• Translational stroke research
• Precision medicine and artificial intelligence

Background:

• Stroke is a complex condition involving multiple biological pathways.
• Single-omics approaches are insufficient; multi-omics data offer deeper insights but pose analytical challenges.
• Machine learning (ML) can address multi-omics complexity, but its predictive accuracy and reproducibility in stroke are under-explored.

Purpose of the Study:

• To systematically review ML models utilizing multi-omics data for stroke risk stratification.
• To analyze discriminatory performance, data integration strategies, and validation/reporting practices.
• To guide future methodological advancements in multi-omics ML for stroke.

Main Methods:

• Systematic literature search (PRISMA 2020) across 9 databases (Jan 2000-Jul 2025).
• Inclusion criteria: adults, stroke prediction, ≥2 omics layers, ML performance metrics reported.
• Risk of bias (PM-ROB) and reporting quality (MIMAR) assessed; primary outcome: Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• 7 studies (n=40,274) met criteria, published 2022-2025, integrating 2 omics layers (e.g., metabolomics-proteomics).
• Supervised ML algorithms included SVMs, tree ensembles, GLMs, and deep learning.
• High apparent discrimination (AUC 0.75-0.97) reported, but only 3 studies performed external validation; calibration and operating points were rarely assessed.

Conclusions:

• Multi-omics ML models demonstrate high apparent stroke risk stratification performance but face methodological limitations.
• Small sample sizes, design heterogeneity, and incomplete reporting impede reproducibility and generalizability.
• Future research requires robust evaluation, external validation, and benchmarking to establish clinical utility for precision stroke care.