Machine learning classification of texture features of MRI breast tumor and peri-tumor of combined pre- and early treatment predicts pathologic complete response

Table 2 ROC metrics for predicting PCR based on molecular subtypes, MRI features at pre- and during NAC using Ensemble RUSBoosted Tree classifier

Time point	Features type	Sens.	Spec.	PPV	NPV	Accuracy	AUC	P-value
–	Molecular subtypes	86.48	76.92	91.42	66.66	84	0.82 (0.66, 0.97)	0.07
Tp1	MRI texture only	86.48	84.62	94.12	68.75	86	0.88 (0.77, 1.0)	0.03
Tp2		97.30	38.46	81.82	83.33	82	0.72 (0.53, 0.91)	0.13
Tp3		92.85	30	78.78	60	76	0.78 (0.62, 0.95)	0.44
Tp1 + Tp2		1.00	76.92	92.50	1.00	84	0.96 (0.92, 1.0)	0.0003
Tp1	MRI texture + molecular subtypes	89.18	92.30	97.06	75.00	90	0.86 (0.75, 0.98)	0.005
Tp2		89.18	69.23	89.18	69.23	84	0.80 (0.64, 0.96)	0.068
Tp3		96.42	50	84.38	83.33	84	0.87 (0.74, 0.99)	0.09
Tp1 + Tp2		94.59	92.31	97.22	85.71	94	0.98 (0.94,1.0)	0.0003

The data were tumor contours based on post-contrast-enhanced MRI with morphological dilation. The numbers in parenthesis show the 95% confidence intervals

ISSN: 1475-925X