Table 3. . Results for the best single-feature (A), two-feature (B) and three-feature (C) prediction models generated from machine-learning algorithms, K-nearest neighbor and naive-Bayes at 24-h post the first radiation treatment, week 1 and 4 of treatment.

A: single-feature classification
Classifier model	Time point	%Sn	%Sp	AUC	%Acc	Best univariate feature
naive-Bayes	24 h	77	83	0.67	80	ΔAAC-CON
	Week 1	85	86	0.77	86	ΔSS-COR
	Week 4	84	85	0.79	85	ΔACE
K-NN	24 h	75	70	0.74	72	ΔAAC-CON
	Week 1	75	85	0.81	81	ΔSAS-ENE
	Week 4	76	79	0.80	77	ΔASD-ENE
B: two-feature classification
Classifier model	Time point	%Sn	%Sp	AUC	%Acc	Best two features
naive-Bayes	24 h	67	73	0.64	70	ΔMBF + ΔAAC-CON
	Week 1	76	84	0.67	80	ΔSS + ΔAAC-COR
	Week 4	75	77	0.75	76	ΔACE + ΔASD
K-NN	24 h	74	78	0.78	76	ΔSS + ΔAAC-CON
	Week 1	73	78	0.77	76	ΔSS + ΔSAS-ENE
	Week 4	76	82	0.81	79	ΔSS-ENE + ΔASD-ENE
C: three-feature classification
Classifier model	Time point	%Sn	%Sp	AUC	%Acc	Best three features
naive-Bayes	24 h	63	69	0.63	66	ΔMBF + ΔSAS-CON + ΔAAC-CON
	Week 1	68	78	0.65	73	ΔSS + ΔSS-COR + ΔAAC-ENE
	Week 4	66	64	0.61	65	ΔMBF + ΔACE + ΔASD
K-NN	24 h	71	76	0.76	77	ΔSS + ΔSI-ENE + ΔAAC-CON
	Week 1	73	81	0.75	77	ΔSS + ΔMBF-ENE + ΔSAS-ENE
	Week 4	79	80	0.82	80	ΔSS-ENE + ΔSI-ENE + ΔASD-ENE

AAC: Average acoustic concentration; Acc: Accuracy; ACE: Attenuation coefficient estimate; ASD: Average scatterer diameter; AUC: Area under curve; CON: Contrast; COR: Correlation; ENE: Energy; HOM: Homogeneity; K-NN: K-nearest neighbor; MBF: Mid-band fit; SAS: Spacing among scatterers; SI: Spectral intercept; S_n: Sensitivity; S_p: Specificity; SS: Spectral slope.