Abstract
The Coronavirus Disease 2019 (COVID-19) is transitioning into the endemic phase. Nonetheless, it is crucial to remain mindful that pandemics related to infectious respiratory diseases (IRDs) can emerge unpredictably. Therefore, we aimed to develop and validate a severity assessment model for IRDs, including COVID-19, influenza, and novel influenza, using CT images on a multi-centre data set. Of the 805 COVID-19 patients collected from a single centre, 649 were used for training and 156 were used for internal validation (D1). Additionally, three external validation sets were obtained from 7 cohorts: 1138 patients with COVID-19 (D2), and 233 patients with influenza and novel influenza (D3). A hybrid model, referred to as Hybrid-DDM, was constructed by combining two deep learning models and a machine learning model. Across datasets D1, D2, and D3, the Hybrid-DDM exhibited significantly improved performance compared to the baseline model. The areas under the receiver operating curves (AUCs) were 0.830 versus 0.767 (p = 0.036) in D1, 0.801 versus 0.753 (p < 0.001) in D2, and 0.774 versus 0.668 (p < 0.001) in D3. This study indicates that the Hybrid-DDM model, trained using COVID-19 patient data, is effective and can also be applicable to patients with other types of viral pneumonia.
Introduction
As COVID-19 transitions from a pandemic to an endemic phase, the medical community remains wary of future outbreaks of Infectious Respiratory Diseases (IRDs). A critical bottleneck during surges is severity assessment: determining which patients require intensive care (ICU) or mechanical ventilation. While Computed Tomography (CT) is a powerful tool for visualizing lung abnormalities like ground-glass opacity (GGO) and consolidation, human interpretation is time-consuming and subjective. Existing Deep Learning (DL) models often act as "black boxes" that might miss subtle pathological patterns or fail to generalize across different types of pneumonia. This research sought to bridge this gap by creating a model that evaluates the lungs from three distinct perspectives: global context, specific lesion characteristics, and quantitative anatomical distribution.
Methodology: The Hybrid-DDM Approach
The researchers proposed a "Hybrid-DDM" (Deep-Deep-Machine) approach. Instead of relying on a single neural network, they ensembled three distinct models to provide a comprehensive assessment:
- $DL_{lung}$ (Global Context): A 3D ResNet-50 trained on "lung-masked" images, allowing the model to see lesions alongside surrounding lung tissue.
- $DL_{lesion}$ (Focused Pathology): A 3D ResNet-50 trained strictly on "lesion-masked" images (GGO and consolidation), forcing the model to focus on the texture and shape of abnormalities.
- $ML_{LIFe}$ (Quantitative Distribution): A Random Forest model trained on 33 Lung Involvement Features (LIFe), representing the proportion of GGO and consolidation across 11 specific lung segments.
To enable this, 2D U-Nets were first utilized to automate the segmentation of the lungs and specific lesions, ensuring the model's inputs were precisely targeted toward pathological areas.
Datasets and Experimental Design
The study utilized a robust multi-center dataset from South Korea, structured into four distinct cohorts to test both internal performance and external generalization:
- D1 (Internal Validation): 805 COVID-19 patients (649 for training, 156 for validation).
- D2 (External COVID-19): 1,138 patients from 7 different cohorts.
- D3 (External Viral): 233 patients with influenza and novel influenza.
- D4 (External Bacterial): 268 patients with bacterial pneumonia (BP).
Key Results and Insights
The Hybrid-DDM model showed statistically significant improvements in the Area Under the Receiver Operating Characteristic curve (AUC) across all viral datasets compared to a baseline 3D ResNet-50:
- Internal COVID-19 (D1): AUC of 0.830 (vs. 0.767 baseline).
- External COVID-19 (D2): AUC of 0.801 (vs. 0.753 baseline).
- Influenza (D3): AUC of 0.774 (vs. 0.668 baseline).
- Bacterial Pneumonia (D4): AUC of 0.709 (vs. 0.704 baseline—no significant improvement).
Generalization and Clinical Relevance
A key finding was the model's ability to generalize. Despite being trained exclusively on COVID-19 data, it performed significantly better than the baseline on influenza (D3). This suggests that viral pneumonias share a common radiological signature—diffuse GGO and specific anatomical distributions—that the hybrid model successfully captured. Conversely, the lack of improvement in bacterial pneumonia (D4) aligns with clinical observations: bacterial infections often present as localized, dense consolidations, differing fundamentally from the viral patterns the model was optimized to detect.
Conclusion
The Hybrid-DDM framework demonstrates that combining deep learning's feature extraction with machine learning's quantitative anatomical analysis provides a superior tool for severity assessment in viral pneumonias. While currently limited by geographic focus and a lack of non-imaging clinical metadata, the model's success in generalizing from COVID-19 to influenza marks a significant step toward universal AI tools for managing future respiratory pandemics.
Summary updated