Feature Importance Guided Network With Missing Value Learning for Pneumonia Pathogen Classification and Biomarker Discovery

Discovering clinical biomarkers through multivariate modeling is crucial for precise pneumonia diagnosis and treatment. However, existing methods often fall short in either their representational capacity or interpretability. Moreover, these methods are highly sensitive to some missing values, leading to a significant performance degradation. In this paper, we propose a Multimodal Feature Importance Guided Network (MFIG-Net) for pneumonia classification and biomarker discovery. MFIG-Net features two novel modules that operate synergistically: a Feature Reconstruction Module (FRM) for handling missing values and a Feature Selection Module (FSM) for identifying biomarkers. The FSM learns coefficients indicative of variable importance, facilitating biomarker discovery and providing real-time feedback to guide the FRM. The dynamic interaction between them enables the FRM to concentrate on learning the most important features by masking and reconstructing less important variables, thereby significantly boosting model performance and robustness. Rigorous validation using over 4,200 subject samples from multiple hospitals demonstrates that our model achieves an AUC score of 95.40%10.41%, significantly outperforming existing methods. Even at 50% data missing rate, it still achieves an AUC score of 93.94%10.35%, while most other methods fall below 70%. Our experiments further demonstrate that MFIG-Net can also extended to other domains, such as multi-omics modeling, thereby highlighting its strong broad applicability. The code is available at: https://github.com/xingzhidaoren/MFIG-Net.