A routine biopsy contains more than a diagnosis: this study asks whether ordinary clinical information and immunohistochemical markers can estimate a patient’s chance of completely responding to chemotherapy—and what that estimate means for prognosis. Pathological complete response, or pCR, is an established surrogate marker for prognosis in patients with breast cancer after neoadjuvant chemotherapy.

Individualized pCR prediction based on clinical information available at biopsy, particularly immunohistochemical markers, may help identify patients who could benefit from preoperative chemotherapy. After neoadjuvant chemotherapy, pCR is a surrogate marker for prognosis in early breast cancer, and it is strongly associated with a favorable prognosis particularly in triple-negative or HER2-positive breast cancer.

Some hormone receptor-positive, HER2-negative patients have an excellent prognosis despite low pCR rates, creating a reason to look beyond response alone. Biomarkers can help identify patients who have a good response to chemotherapy and help justify treatment when the presence or absence of pCR has a substantial impact on survival.

The study focuses on HER2-negative disease because, for HER2-positive patients, the indication for chemotherapy is usually established independently of biomarker values. Combined information allows individualized prediction of pCR that goes beyond the single dichotomized biomarker approach currently used.

The objective was to estimate a patient’s likelihood of pCR from clinical predictors and the immunohistochemical biomarkers ER, PgR, and Ki-67, either as assessed during routine clinical work or using categorical thresholds. The most precise tool was validated in several independent external samples.

This retrospective, single-center, hospital-based observational study included 1166 patients with HER2-negative breast cancer from the ERNEST-B study who underwent neoadjuvant chemotherapy from 2002 to 2020. The patients were at least eighteen years of age and had available information about pCR status; patients with metastases, contralateral breast cancer at primary diagnosis, or incomplete biomarker information were excluded.

Approval for the analyses was obtained from the ethics committee of the University of Erlangen–Nuremberg. pCR was defined as a complete absence of tumor cells from the breast and lymph nodes after chemotherapy at the time of surgery. The basic regression model used established predictors for pCR: age at diagnosis, inverse body mass index, tumor stage, grade, lymph node status, and tumor type.

Age and inverse body mass index were used as cubic spline functions, with their degrees of freedom determined using the Akaike information criterion. Missing predictor values were substituted by predicted expected values based on nonmissing data.

Other prediction models extended the basic model by including the biomarkers of interest. Model performance was assessed for calibration and discrimination using mean squared error, the receiver operating characteristic curve, and the area under the ROC curve.

These measures were obtained by threefold cross-validation with one hundred repetitions to obtain stable and realistic results. Model-building steps used training data, while performance was assessed using validation data that had not been used for model building.

Apparent measures on the complete dataset were calculated to assess overfitting. The model with the smallest cross-validated mean squared error was considered the final model, and it had the best use of ER, PgR, and Ki-67 compared with the other biomarker usages.

The final model was fitted on the complete dataset and analyzed in greater detail. An explicit formula for predicted pCR probability was derived, implemented in an online calculator, and presented graphically as a nomogram.

The IHC3 plus C score incorporates ER, PgR, and Ki-67 with clinical predictors and provides prognostic information on the risk of distant recurrence. The Hannover Breast Cancer Study and two randomized clinical trials, GeparSepto and GeparOcto, were each used to validate the final prediction model fitted on the primary study dataset.

Discrimination ability was assessed using the area under the ROC curve, with its ninety-five percent confidence interval estimated using ten thousand bootstrap samples. Calibration was checked using a calibration plot and a logistic regression model with observed pCR as the outcome and the logit of predicted pCR probability as the only predictor.

In the primary study cohort, 273 of 1166 patients, or 23.4 percent, achieved a pCR. In the validation cohorts, pCR rates ranged from 17.1 percent to 32.7 percent. Patient characteristics grouped by pCR status are shown in Table 1, while biomarker distributions in the primary cohort are shown in the supplementary figure.

Table 2 compares pCR prediction models using apparent and cross-validated measures: AIC, mean squared error, or MSE, and area under the curve, or AUC. The linear biomarker model using ER, PgR, and Ki-67 reports a cross-validated MSE of 0.1336 and AUC of 0.836, while established biomarker categories report 0.1392 and 0.814.

The apparent AUC of 0.845 versus cross-validated AUC of 0.836 also indicates a small amount of overfitting. The prediction model with linear use of ER, PgR, and Ki-67, measured from zero to one hundred percent of positively stained cells, was most accurate, with a cross-validated mean squared error of 0.1336.

Other models with continuous biomarker use were more accurate than the model with established biomarker categories, whose mean squared error was 0.1392, and the model with newly determined cutoff values, whose mean squared error was 0.1356. The mean squared error results were confirmed by the Akaike information criterion and area under the ROC curve statistics.

Table 3 presents the final logistic regression model for predicting pathological complete response, or pCR, with coefficients, odds ratios, confidence intervals, and Wald-test p values. The model includes clinical features and biomarkers: tumor stage has an odds ratio of zero point six two per stage, while estrogen receptor expression has an odds ratio of zero point nine nine per percent and a p value below zero point zero zero zero one.

The authors provide these estimates as the basis for calculating an individual patient’s predicted pCR probability. A high predicted pCR probability indicated the efficacy of chemotherapy and corresponded with the impact of pCR status on prognosis.

The greater the predicted pCR probability, the larger the benefit of a pCR in relation to disease-free survival. The hazard ratio for pCR presence versus absence improved from 0.55, with a ninety-five percent central range of 0.07 to 1.77, at zero percent predicted pCR probability to 0.20, with a range of 0.11 to 0.31, at fifty percent.

Figure two shows how the disease-free survival hazard ratio for patients with a pathological complete response, compared with those without one, changes across predicted pCR probability. The solid curve is the mean from twenty thousand random sample splittings, while dashed curves mark the pointwise ninety-five percent central range.

The authors report hazard ratios improving from zero point five five at zero percent predicted pCR probability to zero point two zero at fifty percent, indicating that pCR status carries greater prognostic impact when predicted pCR probability is higher. In the Hannover Breast Cancer Study, the final prediction model had an area under the ROC curve of 0.827, with a ninety-five percent confidence interval from 0.779 to 0.871.

Sensitivities and specificities were similar to those in the primary cohort, and the model was well calibrated. The calibration intercept and slope did not significantly differ from zero and one, respectively, so updating of the model was not necessary.

Across the three analyzed treatment arms of GeparSepto and GeparOcto, the area under the ROC curve was between 0.754 and 0.795. Predicted pCR probability was satisfactorily precise in the GeparOcto-ETC arm, among patients with low predicted probability in the GeparSepto-paclitaxel arm, and among patients with high predicted probability in the GeparSepto-NAB-paclitaxel arm.

For other patients in the GeparSepto arms, predicted pCR probability was systematically too high or too low, suggesting that the model should be recalibrated before application to GeparSepto-like treatment data. The online calculator and nomogram provide updated predicted pCR probabilities for such patients.

Figure three uses Kaplan–Meier curves to show five-year disease-free survival across pCR-probability classes and, within those classes, IHC3-plus-C score groups. In the complete cohort, the reported five-year DFS rates were zero point seven seven for low pCR probability, zero point six six for intermediate, and zero point seven seven for high.

Panels b through d show how IHC3-plus-C further separates patients within each pCR-probability group, while noting that some small groups were omitted. Differences in performance between the primary and validation cohorts are unavoidable and were expected.

More external validation studies covering different geographic regions, time periods, and healthcare institutions may clarify these differences and possibly improve the prediction model. Adding novel biomarkers, such as tumor-infiltrating lymphocytes, might also improve accuracy.

Prediction in rare tumor types such as lobular breast cancer could be improved by using additional, sufficiently large specific cohorts. The study provides a prediction model for pCR after neoadjuvant chemotherapy using clinical parameters and continuous immunohistochemical markers.

The model’s good performance suggests that therapy decisions could be based on predicted pCR probabilities. Considering both the likelihood of pCR and prognosis identifies patients who may benefit from neoadjuvant chemotherapy and patients with an unexpectedly unfavorable prognosis.

Novel treatment strategies may be necessary for patients with an unexpectedly unfavorable prognosis. A model using clinical parameters and continuous ER, PgR, and Ki-67 markers predicted pathological complete response and retained useful performance in external cohorts.

Combining response probability with recurrence risk could help distinguish patients likely to benefit from neoadjuvant chemotherapy from those needing alternative strategies.