Towards a Biological Evaluation Framework for Oversampling (BEFO) gene expression data.

Area of Science:

• Biomedical research
• Machine learning applications
• Data science

Background:

• Machine learning (ML) models are increasingly used in biomedical research for improved diagnostics and prognostics.
• Biomedical datasets often exhibit class imbalance, leading to biased ML models.
• Existing oversampling techniques lack biological validation for synthetic data, limiting clinical applicability.

Purpose of the Study:

• To introduce the Biological Evaluation Framework for Oversampling (BEFO) to ensure synthetic gene expression data accurately reflects biological patterns.
• To mitigate bias in ML models and enhance the trustworthiness of predictions in clinical settings.
• To establish a new standard for evaluating synthetic data in biomedical ML.

Main Methods:

• Developed a ranking method for synthetic samples based on Weighted Gene Co-expression Network Analysis (WGCNA) gene co-expression clusters.
• Constructed random forests to assess synthetic sample alignment with biological clusters.
• Included only synthetic samples demonstrating higher importance than real samples.

Main Results:

• The BEFO framework improved the biological feasibility of oversampled datasets by an average of 11%.
• Classification performance improved by an average of 9% compared to state-of-the-art methods.
• Evaluated across six real-world gene expression datasets using ten classification algorithms.

Conclusions:

• The proposed ML oversampling framework enhances biological relevance and predictive performance.
• BEFO offers a robust method for validating synthetic data in biomedical ML.
• This approach improves the reliability of ML decision support systems in clinical practice.