Leveraging Scheme for Cross-Study Microbiome Machine Learning Prediction and Feature Evaluations.

Area of Science:

• Microbiome research
• Machine learning applications in medicine
• Disease prediction modeling

Background:

• The human microbiome plays a critical role in various diseases.
• Microbiome data is increasingly utilized for disease prediction.
• Existing models often lack generalizability across independent microbiome studies due to high data variability.

Purpose of the Study:

• To develop a method for enhancing the generalizability and interpretability of machine learning models for disease prediction using microbiome data.
• To address the challenge of applying models trained on one microbiome dataset to others.
• To improve the accuracy of predicting colorectal cancer, Crohn's disease, and immunotherapy response.

Main Methods:

• Developed a data leveraging scheme by combining smaller target datasets with larger source datasets.
• Investigated the impact of source data proportion (minimum 25% of target samples) on model performance.
• Utilized random forest as the primary machine learning model.
• Employed feature selection techniques to identify key microbial taxa.

Main Results:

• Combining datasets, with at least 25% of target samples in the source data, improved model performance.
• Random forest demonstrated superior performance among tested models.
• Feature selection identified common and important taxa predictive of diseases across studies.
• The proposed method enhanced both accuracy and interpretability of microbiome-based disease prediction models.

Conclusions:

• The developed data leveraging scheme is a promising approach for improving microbiome-based machine learning models.
• Enhanced generalizability and interpretability are achievable through strategic dataset combination.
• This method offers a pathway for more robust and reliable disease prediction using microbiome data.