Prediction of pathological complete response after neoadjuvant chemotherapy for HER2-negative breast cancer patients with routine immunohistochemical markers

Table 2 Overall performance of the prediction models for pCR

Prediction model*	Apparent measure^a			Cross-validated measure^b
Prediction model*	AIC	MSE	AUC	MSE	AUC
Null model	1271.1	0.1793	0.500	0.1797 (0.0094)	0.500 (0.000)
Basic model	1100.8	0.1525	0.759	0.1556 (0.0084)	0.748 (0.021)
Basic + ER, PgR, Ki-67 linear^c	942.1	0.1292	0.845	0.1336 (0.0079)	0.836 (0.017)
Basic + ER, PgR, Ki-67 cubic spline, 2 df^c	946.4	0.1291	0.846	0.1343 (0.0078)	0.833 (0.017)
Basic + ER, PgR, Ki-67 cubic spline, 3 df^c	952.2	0.1290	0.846	0.1346 (0.0078)	0.832 (0.017)
Basic + ER, PgR, Ki-67 established categories	991.5	0.1350	0.825	0.1392 (0.0084)	0.814 (0.019)
Basic + ER, PgR, Ki-67 new categories^d	948.1	0.1294	0.844	0.1356 (0.0080)	0.828 (0.018)

AIC, Akaike information criterion; AUC, area under the curve; df, degrees of freedom; ER, estrogen receptor (expression); MSE, mean squared error; pCR, pathological complete response; PgR, progesterone receptor (expression)
^*The null model did not contain any predictors. The basic model included age at diagnosis, body mass index, tumor stage, grade, lymph node status, and tumor type. All other models were extensions of the basic model
^aThe models were fitted on the complete dataset. Confidence intervals were not calculated because model building and application to the same dataset might result in overoptimistic measures. A 95% confidence interval for the AUC would not then cover the true AUC with a 95% likelihood
^bSummary statistics (mean and standard deviation in brackets) of MSE and AUC were obtained by threefold cross-validation with 100 repetitions
^cOnly three out of 27 continuous biomarker models are shown. These are the models with 1 (i.e. linear), 2, or 3 degrees of freedom for all three biomarkers, representing different levels of complexity
^dThe AIC is not a reliable measure for this model because it was applied to data that had already been used to identify the cutoff points

ISSN: 1465-542X