A1284
Title: Improving microbiome modelling and prediction by using an optimised and class-weighted classifier in imbalanced datasets
Authors: Chengxin Li - Leeds Inst of Medical Research (LIMR) (United Kingdom) [presenting]
Henry Wood - (United Kingdom)
Arief Gusnanto - University of Leeds (United Kingdom)
Rishabh Bezbaruah - University of Leeds (United Kingdom)
Abstract: Motivation: Distinct gut microbiome profiles between colorectal cancer (CRC) patients and healthy individuals have enabled the application of machine learning (ML) approaches to microbiome-based colorectal cancer classification. However, the limited transferability of models across cohorts remains a major challenge, as severe class imbalance and heterogeneous analysis pipelines can lead to unstable performance and reduced clinical utility. Results: The impact of normalisation, taxonomic resolution, and class imbalance handling on microbiome-based CRC prediction was evaluated. Using a real-world screening cohort, Random Forest and XGBoost classifiers were developed and evaluated. GMPR normalisation consistently improved performance across models. At the genus level, including uncultured taxa, XGBoost with class weighting achieved the highest performance and outperformed Random Forest. Both models exceeded the performance of previously published approaches on the same cohort. External validation showed higher mean AUCs compared with prior studies, although XGBoost exhibited reduced generalisability across more dissimilar cohorts.
