With the importance of artificial intelligence emerging in the medical field, predictive technology using multi-omix data and machine learning is expected to be an innovative future medical technology for cancer diagnosis and treatment. However, predictive factors and models vary by researchers, centers and analysis platforms, requiring more objective and reliable model discovery and verification. Furthermore, modeling through machine learning is often difficult to interpret, so not only the accuracy of predictions but also the interpretability of models are recognized as critical factors in evaluating their performance.
Korean researchers have developed a new type of cancer diagnosis artificial intelligence platform with researchers at the Memorial Sloan Kettering Cancer Center in the U.S. The Korea Research Foundation said Professor Kim Sung-young’s research team at Konkuk University has developed an artificial intelligence platform that can distinguish cancer with high reliability using meta-analysis-based machine learning algorithms.
Meta-analysis in the medical field is a technique that evaluates the consistency of results and increases statistical accuracy by integrating various research results on the same subject. It is a key statistical technique in evidence-driven medicine by converting research estimates from individual studies conducted under similar themes to a common effect size to objectively generalize experimental results.
Konkuk University’s research team integrated gene expression and related biological pathways using meta-analysis-based algorithms and used them as learning materials for artificial intelligence.
The research team combined machine learning algorithms with meta-analysis to build more robust models (MLMA). The research team confirmed that the dimension reduction method using nonlinear principal component analysis and biological path greatly increases the generality and interpretability of the model.
The team validated these algorithms on real-world cancer-blind thyroid cancer samples which showed near-perfect classification performance. Several subtypes of thyroid cancer were tested and even these subtypes were distinguished with high accuracy.
The research team maximized the interpretability of the model by finding control factors of the biological path related to thyroid cancer through multi-omix analysis.
The research team said that the algorithm can be easily applied to other cancers in addition to thyroid cancer. It is also expected to be preferred in clinicians who value the consistency of cause analysis and results with regression and path analysis-based algorithms.
The study will be published in the latest issue of the “Briefings in Bioinformatics.”