Abstract
Machine learning has been increasingly applied to microbiome data for biomarker discovery. However, microbiome datasets are typically high-dimensional, sparse, and correlated, which makes model training challenging and prone to overfitting. Previous studies have also reported that microbiome features exhibit binary-like characteristics, and that binarization does not necessarily reduce predictive performance. This observation motivated our work. Building on this idea, we propose a SHAP-based binarization pipeline. We first trained several machine learning models on raw continuous data and selected the best-performing model (random forest). Using SHAP values derived from the training set, we determined feature-specific thresholds that best separated positive and negative contributions. The dataset was then binarized using these thresholds and new models were trained on the transformed data. We evaluated this approach on gut microbiome abundance data (283 species, 220 genera, 1,569 individuals) to classify inflammatory bowel disease (IBD) versus healthy controls. The SHAP-based binarization consistently improved classification performance and interpretability compared with both continuous data and zero-threshold binarization. The best model's Matthews correlation coefficient increased from 0.884 to 0.928, with the largest improvements observed in non-tree-based models such as logistic regression and neural networks. SHAP summary plots also revealed clearer feature patterns, and biomarker rankings were more stable. In addition, the pipeline enabled us to identify a concise set of 17 microbial biomarkers associated with IBD. This study introduces a novel approach for microbiome data analysis by explicitly linking binarization thresholds to SHAP-derived feature contributions. Our approach was grounded in the observation of binary-like patterns revealed through SHAP values. Furthermore, although binarization inevitably raises concerns about information loss, our evaluation confirmed improvements not only in predictive performance but also in interpretability and biomarker stability, providing a broader validation of robustness. These findings highlight SHAP-based binarization as an effective strategy for high-dimensional microbiome data, with broad applicability and opportunities for future extension.