A Method to Extract Feature Variables Contributed in Nonlinear Machine Learning Prediction.

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• Advanced machine learning models like deep neural networks offer high prediction accuracy but lack interpretability.
• The
• black box
• nature of these models hinders trust and reliability in their predictions.
• Improving the explanatory power of nonlinear machine learning models is essential for their clinical adoption.

Purpose of the Study:

• To develop a novel factor analysis technique for presenting feature variables in nonlinear machine learning models.
• To enhance the interpretability and reliability of predictions made by complex AI models.
• To enable the identification of key predictive factors in biological and medical data.

Main Methods:

• A two-part factor analysis technique was developed, comprising backward analysis and factor extraction.
• Factor extraction identifies feature variables from the posterior probability distribution calculated via backward analysis.
• The technique was applied to gene expression data from prostate cancer patients.

Main Results:

• Deep neural networks achieved approximately 5% higher prediction accuracy than support vector machines on prostate tumor gene expression data.
• The developed factor extraction method showed moderate concordance (40-49%) with existing methods for identifying top feature variables.
• The technique demonstrated the potential for evaluating machine learning models from different perspectives.

Conclusions:

• The proposed factor analysis technique offers a new approach to understanding feature importance in nonlinear machine learning models.
• This method can improve the reliability and trustworthiness of AI-driven predictions in scientific research.
• Future work includes validating extracted features and applying them to clinical studies.