Table 3.

Ablations on the intermediate model.

Model	AUC	Accuracy	Specificity	Sensitivity
Final component architecture
FC layer	0.9* [0.82–0.95]	0.89	0.9	0.8
1DCNN	0.91* [0.82–0.95]	0.89	0.89	0.9
XGBoost	0.92* [0.84–0.97]	0.88	0.88	0.9
Dimensionality reduction
w/o	0.92* [0.87–0.97]	0.85	0.84	0.9
Sparse PCA	0.94 [0.88–0.98]	0.89	0.89	0.9
NMF	0.94 [0.88–0.98]	0.85	0.84	0.9
Modified LLE	0.96 [0.92–1.0]	0.9	0.92	0.8
Spectral embedding	0.93* [0.89–0.97]	0.85	0.84	0.9
BiAttention	0.95 [0.9–0.99]	0.9	0.9	0.9
Embedding dimension
d = 1024	0.95 [0.9–0.98]	0.89	0.89	0.9
d = 256	0.96 [0.93–1.0]	0.92	0.92	0.9
Full	$\mathbf{0.96}[\mathbf{0.93}-\mathbf{1.0}]$	$\mathbf{0.93}$	$\mathbf{0.94}$	$\mathbf{0.9}$

Asterisks mark statistically significant difference compared to the full intermediate model architecture (Kolmogorov-Smirnov test with $p\le 0.05$ ).

Best performing values are in [bold].