Table 7.
Robustness analysis aims to examine whether our framework can provide COVID-19 decision support for new regions.
| Training regions | Methods | Year | Testing Regions | |
|---|---|---|---|---|
| Spain | China | |||
| Spain (BIMCV-COVID-19) | CLIP54 | 2021 | 85.4 | 75.4 |
| ConVIRT34 | 2022 | 92.7 | 85.6 | |
| BioViL30 | 2022 | 92.8 | 83.3 | |
| Med-MLLM | Ours | 95.2 | 90.1 | |
| China (COVID-HCH) | CLIP54 | 2021 | 70.5 | 80.7 |
| ConVIRT34 | 2022 | 79.9 | 91.7 | |
| BioViL30 | 2022 | 77.0 | 90.4 | |
| Med-MLLM | Ours | 84.8 | 93.9 | |
| >20 Countries (COVID-CXR | CLIP54 | 2021 | 63.3 | 61.5 |
| excl. Spain & China) | ConVIRT34 | 2022 | 71.2 | 69.0 |
| BioViL30 | 2022 | 69.6 | 70.4 | |
| Med-MLLM | Ours | 78.2 | 74.8 | |
We perform cross-region prediction by training on patient data from one region and evaluating on patient data with different phenotypes from other regions. All values are reported in percentage (%). The best results are in bold.