Authors:

Rebekah E Mou

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Antimicrobial resistance (AMR), a major modern healthcare problem, can be limited by supplementing heavy antibiotic regimes with new treatments called fecal microbiota transplants (FMTs) in the case of patients for whom antibiotics have become ineffective. Machine learning (ML) methods have emerged as a potentially fast and effective way to visualize and predict AMR gene dynamics, though there have been few direct comparisons between different machine learning methods within this application. This research aims to establish direct comparisons between different machine learning methods in analyzing or predicting frequencies of antimicrobial resistance genes in post-FMT patients. To do this, we trained various supervised ML methods (Random Forest, XGBoost, and Linear Regression/Lasso) to predict post-FMT AMR gene frequencies. Additionally, unsupervised ML models like PCA and t-SNE were used for exploratory analysis and visualization of gene clusters. Using the hyperparameter optimization framework Optuna, we determined the optimal hyperparameters for each ML model and evaluated the most accurate supervised ML method for each gene class. For both supervised and unsupervised models, SHapley Additive exPlanations (SHAP) was used to determine feature importance and model interpretability. Preliminary results suggest that Random Forest, while commonly regarded as a strong baseline in machine learning, performed less accurately than other models in predicting AMR gene abundance. This research aims to provide groundwork for further comparison of ML methods in an AMR space and background for the usage of the most optimal ML methods in predicting AMR gene frequencies. Keywords: Antimicrobial Resistance; Artificial Intelligence; Machine Learning; Microbiome; Fecal Microbiota Transplant

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Rebekah E Mou