Improving platelet-RNA-based diagnostics: a comparative analysis of machine learning models for cancer detection and...

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Liquid biopsy offers a minimally invasive, cost-effective method for cancer detection and monitoring.
• Machine learning (ML) is crucial for analyzing complex liquid biopsy data, particularly RNA expression profiles.
• Current ML methods lack standardization for liquid biopsy data analysis.

Purpose of the Study:

• To evaluate and compare various ML techniques for pan-cancer detection and multiclass classification using platelet RNA.
• To establish a standardized, high-performing ML pipeline for liquid biopsy data analysis.
• To provide an open-source baseline for future research in cancer diagnostics.

Main Methods:

• A large-scale study analyzed platelet RNA samples from 1397 cancer patients (17 types) and 354 healthy donors.
• Data quality was ensured through rigorous patient data filtering.
• Multiple ML models, including feature selection for RNA transcripts, were assessed for performance.

Main Results:

• Simple logistic regression achieved the highest cancer detection rate (68%) with 99% specificity.
• Multiclass classification accuracy reached 79.38% for distinguishing between five cancer types.
• The study's ML approach improved cancer detection by 5% and multiclass classification by 9.65% over previous methods.

Conclusions:

• Classical ML models, like logistic regression, are highly effective for platelet RNA-based cancer diagnostics.
• The developed methods provide a significant improvement over existing techniques.
• Open-sourcing code and data pipelines will facilitate further research and development in the field.