TABLE 5.

The comparison performance of machine learning models using qNet feature, ${\text{APD}}_{\mathbf{50}}$ & ${\text{Ca}}_{\text{diastole}}$ features, and 14 in silico features in predicting high, intermediate, and no TdP risk of drugs.

Model	Features	Accuracy	Sensitivity	Specificity	AUC
KNN	qNet	0.608 (0.606–0.609)	0.403 (0.401–0.406)	0.702 (0.701–0.704)	0.562 (0.559–0.565)
	${\text{APD}}_{50}$ & ${\text{Ca}}_{\text{diastole}}$	0.747 (0.741–0.753)	0.582 (0.576–0.588)	0.809 (0.806–0.812)	0.758 (0.746–0.769)
	14 In Silico Features	0.762 (0.761–0.763)	0.607 (0.606–0.608)	0.816 (0.815–0.817)	0.711 (0.710–0.712)
XGBoost	qNet	0.718 (0.717–0.719)	0.515 (0.514–0.516)	0.775 (0.774–0.776)	0.702 (0.701–0.703)
	${\text{APD}}_{50}$ & ${\text{Ca}}_{\text{diastole}}$	0.789 (0.784–0.796)	0.627 (0.621–0.632)	0.804 (0.834–0.841)	0.804 (0.799–0.809)
	14 In Silico Features	0.798 (0.785–0.810)	0.650 (0.620–0.679)	0.841 (0.829–0.854)	0.883 (0.867–0.899)
RF	qNet	0.718 (0.717–0.719)	0.515 (0.514–0.516)	0.775 (0.774–0.776)	0.703 (0.702–0.70)
	${\text{APD}}_{50}$ & ${\text{Ca}}_{\text{diastole}}$	0.790 (0.789–0.791)	0.616 (0.615–0.617)	0.834 (0.833–0.835)	0.794 (0.793–0.795)
	14 In Silico Features	0.835 (0.831–0.839)	0.705 (0.697–0.714)	0.869 (0.866–0.872)	0.905 (0.904–0.906)
ANN	qNet	0.719 (0.718–0.720)	0.517 (0.515–0.519)	0.776 (0.775–0.777)	0.705 (0.704–0.706)
	${\text{APD}}_{50}$ & ${\text{Ca}}_{\text{diastole}}$	0.790 (0.788–0.791)	0.620 (0.616–0.625)	0.835 (0.833–0.836)	0.801 (0.799–0.803)
	14 In Silico Features	0.852 (0.846–0.857)	0.762 (0.758–0.766)	0.888 (0.886–0.889)	0.911 (0.907–0.914)

The bold values mean the highest performance obtained.