Improving Microbiome-Based Disease Prediction With SuperTML and Data Augmentation

The efficacy of feedforward neural networks in analyzing high-dimensional tabular data is limited, especially in the presence of data scarcity, a substantial challenge in microbiome research. Limited data availability can lead to overfitting and poor predictive performance, reducing the applicability of classical neural network models. To address these challenges, this study evaluates SuperTML, a novel deep learning framework designed for small tabular datasets, in the context of microbiome-disease prediction, an unexplored domain for this approach. We compare SuperTML’s performance with classical neural networks and the state-of-the-art DeepMicro framework. Our results show that SuperTML, especially when combined with data augmentation techniques, achieves competitive performance, outperforming classical approaches in five out of six datasets. This study also highlights the importance of data augmentation in boosting model performance and suggests SuperTML as a valid alternative for microbiome-based disease prediction. Finally, we discuss the framework’s limitations, particularly those related to data dimensionality, computational requirements and interpretability, and provide insights into possible solutions for future applications.