Table 11.

Ablation study results for classification performance.

Methods	External dataset 1 (CNUH)			External dataset 2 (YNU)			External dataset 3 (KNUH)
	AUC (95% CI)			AUC (95% CI)			AUC (95% CI)
	Normal	Others	COVID-19	Normal	Others	COVID-19	Normal	Others	COVID-19
Proposed model	0.968 (0.954–0.981)	0.926 (0.893–0.959)	0.953 (0.935–0.971)	0.973 (0.964–0.983)	0.935 (0.914–0.955)	0.884 (0.861–0.906)	0.961 (0.950-0.972)	0.861 (0.837–0.885)	0.898 (0.878–0.918)
w/o pre-train	0.898${}^{**}$ (0.868–0.927)	0.869${}^{*}$ (0.815–0.922)	0.848${}^{**}$ (0.806–0.891)	0.962 (0.950-0.975)	0.919 (0.896–0.942)	0.895 (0.873–0.917)	0.887${}^{**}$ (0.868–0.907)	0.811${}^{**}$ (0.784–0.837)	0.812${}^{**}$ (0.783–0.841)
w/o Transformer	0.935${}^{**}$ (0.914–0.957)	0.893${}^{*}$ (0.846–0.939)	0.914${}^{*}$ (0.887–0.942)	0.964${}^{*}$ (0.953–0.975)	0.899${}^{**}$ (0.875–0.924)	0.846${}^{**}$ (0.820-0.872)	0.925${}^{**}$ (0.910-0.940)	0.791${}^{**}$ (0.762–0.820)	0.794${}^{**}$ (0.765–0.822)
w/o PCAM	0.943${}^{*}$ (0.921–0.965)	0.874${}^{*}$ (0.817–0.931)	0.911${}^{*}$ (0.882–0.941)	0.968 (0.957–0.979)	0.911${}^{*}$ (0.886–0.937)	0.892 (0.870-0.915)	0.938${}^{**}$ (0.924–0.952)	0.868 (0.845–0.890)	0.817${}^{**}$ (0.788–0.846)
w/o position embedding	0.965 (0.950-0.980)	0.939 (0.909–0.969)	0.940${}^{*}$ (0.918–0.963)	0.976 (0.967–0.985)	0.925${}^{*}$ (0.902–0.947)	0.893${}^{†}$(0.871–0.914)	0.963 (0.952–0.973)	0.861 (0.838–0.885)	0.878${}^{**}$ (0.856–0.901)

Note:${}^{*}$, ${}^{**}$ denote the better performance of our model, while ${}^{†}$, ${}^{††}$ denotes worse performance of our model with statistical significance ($p<0.05$, $p<0.001$). CI: confidence interval.