Intestinal Flora Disease Prediction Based on Global Hybrid Dilated Convolution

Introduction: With the development of 16S rRNA gene amplification and sequencing technology, studies have found that the relative abundance of intestinal microorganisms impacts the diagnosis of diseases. Data analysis techniques have been widely used in the field of bioinformatics. However, the traditional algorithms have a fixed structure, and they have reached a bottleneck. The fully connected neural network assumes that there is a connection between each bacterial group. Common convolutional neural networks (CNNs) are suitable for image tasks where the pixels in the feature area are closely arranged. However, the exchange invariance of rows and columns of the operational taxonomic unit (OTU) table makes the algorithm's performance less than ideal. Methods: In this study, we have improved the data preprocessing technology and proposed a global hybrid dilated convolution framework, using multiple hybrid dilated convolution encoders. We then predicted the disease through multilayer perceptron (MLP) decoder and logistic regression. Results: We used various indicators to evaluate the algorithm and enhance the transparency through Shapley additive explanations (SHAP). Our method achieved state-of-the-art (SOTA) performance and demonstrated significant improvement across 4 datasets on all the indicators. Its feature extraction ability aligned with clinical studies employing SHAP analysis. Discussion: Cross-dataset prediction presents challenges in terms of data heterogeneity and technical variability. Future efforts can prioritize harmonized taxonomic standards and adversarial training. The proposed framework has demonstrated robustness to tabular data permutations. Conclusion: The global hybrid dilated convolution framework can effectively perform disease prediction through solely OTU tables. It can alleviate the crucial issues of former methods and can aid in related clinical microbiome or tabular data analysis.