Benchmarking ensemble machine learning algorithms for multi-class, multi-omics data integration in clinical outcome...

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Multi-modal patient data offers complementary insights for disease modeling.
• Analyzing multi-modal, multi-omics data presents significant computational challenges.
• Effective integration strategies are crucial for leveraging diverse data types.

Purpose of the Study:

• To compare ensemble machine learning (ML) algorithms for late integration of multi-class patient data.
• To evaluate the performance of various ensemble methods including voting, meta-learners, and boosted models.
• To assess the efficacy of these methods on real-world datasets for hepatocellular carcinoma, breast cancer, and irritable bowel disease.

Main Methods:

• Comparison of ensemble ML algorithms: voting (hard/soft), meta-learner, multi-modal AdaBoost (hard/soft/meta-learner), PB-MVBoost, and mixture of experts.
• Late integration of multi-class data from different modalities.
• Evaluation using in-house hepatocellular carcinoma data and external validation datasets for breast cancer and irritable bowel disease.

Main Results:

• Ensemble ML models achieved an area under the receiver operating curve (AUC) of up to 0.85.
• PB-MVBoost and AdaBoost with soft vote demonstrated the best performance among the tested methods.
• Integrated models produced more stable predictive features and enhanced accuracy compared to individual modalities or simple concatenation.

Conclusions:

• Integrating complementary omics and data modalities using effective ensemble ML models significantly enhances accuracy in multi-class clinical outcome predictions.
• Boosted ensemble methods, particularly PB-MVBoost and AdaBoost with soft vote, are highly effective for multi-modal data integration.
• The study provides recommendations for optimal integration of multi-modal, multi-class data in clinical research.