Machine Learning-Driven Classification of Type 2 Diabetes Using Gut Microbiome Profiles for Enhanced Detection and Personalised Therapeutics

The purpose of the study is to investigate Type 2 Diabetes Mellitus (T2DM) progression and risk through investigating gut microbiome biomarkers, providing insights into disease status and potential synbiotic and fecal microbiota transplantation interventions. Gut dysbiosis serves as a key indicator for T2DM detection, supporting effective intervention strategies by outlining potential targets for up/down-regulation. Our preliminary research aimed to compile gut microbiome biomarkers documented to be associated with disease states, forming the foundation for constructing a model capable of disease state analysis. Gut microbiome composition data of individual patients from sequenced metagenomic files were compiled, encompassing the abundance of over 4,500 microbial taxa across profiles. Multiple models were tested to evaluate the greatest efficacy given the nature of the data used, from which XGBoost was selected and optimized. Despite 95.8% sparsity in our microbiome data alongside high local intrinsic dimensionality (LID{approx}12), an 80% testing accuracy was achieved. SHAP values enabled the analysis of key bacterial features, alongside a correlational network for biological validation through the visualization of co-occurring T2DM-associated bacterial taxa. This supported the hypothesis that machine learning can provide critical insights into the nature of dysbiosis, precisely classifying disease states and modelling biologically-accurate relationships. Our findings demonstrate the biological accuracy of the model, along with its ability to identify under-researched species significantly correlated with disease states. This suggested their potential as stable and reliable biomarkers. While further data is required to enhance model performance, this work represents a crucial step toward revolutionizing targeted diagnosis, patient care, and intervention strategies.