Dynamic ultrasound-based modeling predictive of response to neoadjuvant chemotherapy in patients with early breast cancer

Abstract

Early prediction of patient responses to neoadjuvant chemotherapy (NACT) is essential for the precision treatment of early breast cancer (EBC). Therefore, this study aims to noninvasively and early predict pathological complete response (pCR). We used dynamic ultrasound (US) imaging changes acquired during NACT, along with clinicopathological features, to create a nomogram and construct a machine learning model. This retrospective study included 304 EBC patients recruited from multiple centers. All enrollees had completed NACT regimens, and underwent US examinations at baseline and at each NACT cycle. We subsequently determined that percentage reduction of tumor maximum diameter from baseline to third cycle of NACT serves to independent predictor for pCR, enabling creation of a nomogram (). Our predictive accuracy further improved () by combining dynamic US data and clinicopathological features in a machine learning model. Such models may offer a means of accurately predicting NACT responses in this setting, helping to individualize patient therapy. Our study may provide additional insights into the US-based response prediction by focusing on the dynamic changes of the tumor in the early and full NACT cycle.

Introduction

Neoadjuvant chemotherapy (NACT) is systemic treatment administered prior to surgery. In patients with early breast cancer (EBC), NACT may mitigate surgical damage and broaden surgical possibilities, improving eligibility for such options as breast-conserving surgery and exempted axillary node dissection¹. Individual NACT regimens may also be adjusted in accord with observed patients’ drug sensitivities². In recent years, indications for NACT have increased substantially^3–6, especially in terms of triple-negative breast cancer (TNBC) and HER2-enriched (HER2+) breast cancer⁷. Patient-level analysis has shown that achieving pathological complete response (pCR) after NACT is associated with increased event-free survival (EFS) and overall survival (OS) compared with patients with non-pCR^2,8–10. However, only 5–38% of patients with EBC actually achieve pCR, with the majority having residual disease, meaning failure of NACT^2,8,9. It is particularly important that therapeutic responses are accurately evaluated, because ineffective NACT will potentially increase resistant heterogeneous clones, inducing drug resistance and compounding toxic side effects. Treatment strategies should thus be revised timely to reduce unnecessary drug burdens and avoid suboptimal surgical timing. Hence, early and accurate identification of tumor response to NACT is critical for precision treatment decision-making and individualization of surgical procedures.

Mammography, ultrasound (US), and magnetic resonance imaging (MRI) are often regularly undertaken in patients receiving NACT for purpose of therapeutic response monitoring. MRI is currently deemed the most accurate approach to detect tumor response to NACT^11–14, but procedural drawbacks (i.e., high cost and discomfort during testing) limit its use in this setting. Mammography is less accurate than US in predicting residual tumor size after NACT^15,16. Liquid biopsy, such as circulating tumor cells (CTCs) and circulating tumor DNA (ctDNA) may readily reflect the efficacy of NACT in early-stage^17,18. Unfortunately, liquid biopsy is clinically problematic, owing to its invasiveness, low content, the insufficient sensitivity of next-generation sequencing technology (NGS), and a lack of adequate medical evidence.

Various benefits of US afforded in non-invasiveness, economy, operability, accessibility, and freedom from radiation makes it the best method for frequent, dynamic monitoring of NACT efficacy. Studies have shown that US is a useful tool for early assessment of pCR after NACT, even though it is less reliable in predicting poor pathological outcomes¹⁹. Although some studies have utilized US and MRI parameters acquired before NACT to predict pCR²⁰, tumor characteristics are apt to change during NACT²¹. It is thus doubtful that a single pre- or post-NACT static state is an accurate representation of patients’ response to NACT. While not entirely foolproof, dynamic changes in tumors during NACT may offer clues for predicting responses to NACT.

In recent years, artificial intelligence has been invoked for tumor epidemiologic analysis, auxiliary tumor imaging diagnostics, and prognostic predictive models. Existing research aimed at machine learning methods has relied on mammography, MRI, and US imaging to predict the efficacy of NACT in patients with breast cancer^22–24. Herein, we have constructed a nomogram and a support vector machine (SVM) for early prediction of pCR, based on both clinicopathological characteristics of EBC patients and longitudinal dynamic changes in US images during NACT. Our intent was to help clinicians make critical treatment decisions and formulate appropriate surgical plans.

Methods

Patient population

This multicenter retrospective study enrolled 506 patients with EBC from four medical centers between October 2013 and December 2021. All had been diagnosed with primary breast cancer and treated with NACT. We excluded 202 of these candidates on the following grounds: no surgery performed (), incomplete NACT (), missing clinicopathological data (), unqualified US report (), less than three US examinations (), male breast cancer (), and inflammatory breast cancer (). Ultimately, 304 patients remained for randomization to training and validation sets at a 7:3 ratio. The study design is shown in Fig. 1. Figure 1 was created by Figdraw (https://www.figdraw.com/static/index.html).

Fig. 1.

Research process (Created by Figdraw). (a) Study enrollment process, randomly assigning 304 patients to a training or validation set at 7:3 ratio. (b) Each of the 304 patients selected received 4–8 cycles of neoadjuvant chemotherapy (NACT), once diagnosed with early breast cancer (EBC) by core needle biopsy. Ultrasound (US) examinations were done before NACT and at each cycle of NACT. Upon completing NACT, surgery was performed, and postoperative pathological test was used to determine the pathological response. (c) Our research objective was two-pronged. The first part was to predict early NACT response through dynamic change in tumor during first three treatment cycles, to gauge forthcoming treatment needs. The second was predicting pCR preoperatively, based on dynamic tumor changes at each cycle of NACT, to formulate operative strategies.

The study was approved by the ethics committee of Second Affiliated Hospital of Dalian Medical University. All methods were performed in accordance with the relevant guidelines and regulations. Informed consent has been obtained from patients.

Clinical and pathological evaluations

We collected patient clinical characteristics at baseline (before NACT), including age and clinical tumor/nodal stage. Data on hormone receptors (estrogen receptor, ER; progesterone receptor, PR) and other markers (HER2 and Ki67) were obtained from pre-treatment core needle biopsies. Hormone receptor positivity (either ER+ or PR+) was defined as ≥ 1% of tumor cells staining positively in immunohistochemistry (IHC)²⁵. Tumors with immunohistochemical HER2 3+ or HER2 2+ with HER2 gene amplification confirmed by fluorescent in situ hybridization (FISH) were considered HER2+²⁶. The Ki67 index was denoted by percentage of stained cells²⁷. In accord with the above criteria, we classified tumors as HR+/HER2+, HR−/HER2+, HR+/HER2−, or HR−/HER2−.

Neoadjuvant chemotherapy and pathological response

Each patient was treated with 4–8 cycles of NACT, based on anthracycline (A) and/or taxanes (T). Some patients received carboplatin (Cb) as well. The patients with HER2+ tumors were treated with trastuzumab or trastuzumab plus pertuzumab in conjunction with chemotherapy.

Biopsy specimens acquired before NACT and surgical specimens assessed after NACT were used to evaluate pathological response, applying the Miller-Payne system²⁸. Pathological complete response (pCR) was indicated by the absence of invasive tumor cells in both breast and axillary lymph nodes, and ductal carcinoma in situ could be present (ypT0/is ypN0)^29–31.

Ultrasound examinations

All US scans of breast were performed by three experienced radiologists independently. All radiologists were required to undergo a unified training in operating technology according to the guidelines of the American College of Radiology (BI-RADS lexicon) and the expert consensus of the National Ultrasound Quality Control Centre. The largest radial section of the lesion (longitudinal section) and the largest section vertical to it (transverse section) were required to be obtained. Two maximum diameters vertical to each other were measured in the longitudinal section of the lesion, and then the transverse diameter was measured in the transverse section. The maximum tumor diameter was selected from the two sections. Volumes were calculated from recorded three maximum diameters. Patients routinely underwent US examinations at baseline and at every cycle of NACT to determine maximum tumor diameters and volumes over time. To measure early dynamic changes of tumors, we defined as , where and were maximum tumor diameters at baseline and at NACT cycle x, respectively.

Nomogram development

Logistic regression was done to identify associations of clinical, pathological, and US parameters with pCR. We reserved variables of significance (p < 0.05) in univariate analysis for multivariate analysis, determining independent predictive factors of pCR. The identified factors provided a basis for nomogram construction. Predictive accuracy of was assessed by the receiver operating characteristic (ROC) curve, deriving sensitivity and specificity from the area under the curve (AUC). Values of AUC (range, 0–1) indicate perfect concordance at 1, no better than chance at 0.5, and discordance at 0³². Internal validation of the nomogram was performed by a calibration method and by the AUC in the training set. External validation was conferred by performing AUC in the validation set. The calibration curve, assessed by Hosmer–Lemeshow goodness-of-fit test (p > 0.05 indicating good fit), and the decision curve analysis (DCA) served to evaluate predictive model utility. The final nomogram was conducted using R software (v4.2.1; The R Foundation for Statistical Computing, Vienna, Austria).

Machine learning model

The samples were randomly divided into a training set and a validation set in a 7:3 ratio. Various parameters, including ER, PR, HER2, and Ki67 status; NACT regimen; and maximum tumor diameters/volumes by cycle enabled the construction of our machine learning model. Data were preprocessed in three steps, the first being observed dynamic tumor changes. For this, we excluded any samplings examined less than three times. At the second step, data were standardized to eliminate differences between features. Considering the imbalance of data types, the pCR sampling was extended as part of the training set for step three, using the Synthetic Minority Over-sampling Technique (SMOTE) method. We then constructed the pCR prediction model using the support vector machine (SVM). SVM is a binary classification model, transforming classification issues into convex quadratic programming for problem solving. The radial basis function (RBF) was used as the kernel in SVM. To avoid overfitting, we conducted fivefold cross-validation when selecting parameters for RBFs. This model was evaluated by precision, recall, and F1-score, all obtained from the confusion matrix. To assess the predictive performance of the model, ROC curves and AUC were calculated. The process was conducted using Python (v3.8.12; Python Software Foundation).

Statistical analysis

We compared patient characteristics of the training and validation sets. Mann–Whitney U test was applied to compare between pCR and non-pCR patients at different NACT cycles. The Hosmer–Lemeshow test was used to assess the fitness of the nomogram. All statistical tests were two-sided, driven by standard softwares (SPSS v25.0 [IBM Corp, Armonk, NY, USA] and R v4.2.1).

Results

Patient characteristics

We randomly assigned 304 EBC patients to a training () or a validation () set. Among all enrollees, 85 (28.0%) achieved pCR. Clinical characteristics of the patient population are summarized in Table 1. More than half (53.0%) of those achieving pCR received anthracycline and taxane agents as chemotherapy, and HER2+ tumors were more often associated with pCR (HR+/HER2+, 28.2%; HR−/HER2+, 33.0%). The pCR rate in HR+/HER2− subtype was 28.2%. Patients with HR−/HER2− subtype were least likely (10.6%) to achieve pCR.

Table 1.

Clinicopathological characteristics of enrolled patients with EBC.

Characteristic	Overall (304)		Training set (213)		Validation set (91)
	pCR (85)	Non-pCR (219)	pCR (52)	Non-pCR (161)	pCR (33)	Non-pCR (58)
Age (year), (n, %)
≤ 50	46 (54.1)	102 (46.6)	27 (51.9)	71 (44.1)	19 (57.6)	31 (53.4)
> 50	39 (45.9)	117 (53.4)	25 (48.1)	90 (55.9)	14 (42.4)	27 (46.6)
cT stage, (n, %)
T1	15 (17.7)	9 (4.1)	8 (15.4)	7 (4.3)	7 (21.2)	2 (3.4)
T2	58 (68.2)	156 (71.2)	34 (65.4)	112 (69.6)	24 (72.7)	44 (75.9)
T3	12 (14.1)	53 (24.2)	10 (19.2)	41 (25.5)	2 (6.1)	12 (20.7)
T4	0 (0)	1 (0.5)	0 (0)	1 (0.6)	0 (0)	0 (0)
cN stage, (n, %)
N0	6 (7.1)	40 (18.2)	3 (5.8)	27 (16.8)	3 (9.1)	13 (22.4)
N1	70 (82.4)	155 (70.8)	42 (80.8)	116 (72.1)	28 (84.8)	39 (67.3)
N2	2 (2.3)	3 (1.4)	2 (3.8)	2 (1.2)	0 (0)	1 (1.7)
N3	7 (8.2)	21 (9.6)	5 (9.6)	16 (9.9)	2 (6.1)	5 (8.6)
T0 (cm), M (P25, P75)	3.5 (2.7, 4.84)	3.5 (2.7, 4.82)	3.735 (2.7, 5.0)	3.785 (2.7, 5.0)	3.2 (2.5, 4.4)	3.2 (2.53, 4.4)
ΔT3 (%), M (P25, P75)	0.26 (0.10, 0.45)	0.26 (0.09, 0.45)	0.26 (0.07, 0.46)	0.26 (0.07, 0.46)	0.27 (0.11, 0.44)	0.26 (0.11, 0.43)
ER status, (n, %)
Positive	37 (43.5)	164 (74.9)	26 (50.0)	122 (75.8)	11 (33.3)	42 (72.4)
Negative	48 (56.5)	55 (25.1)	26 (50.0)	39 (24.2)	22 (66.7)	16 (27.6)
PR status, (n, %)
Positive	42 (49.4)	156 (71.2)	28 (53.8)	118 (73.3)	14 (42.4)	38 (65.5)
Negative	43 (50.6)	63 (28.8)	24 (46.2)	43 (26.7)	19 (57.6)	20 (34.5)
HER2 status, (n, %)
Positive	52 (61.2)	60 (27.4)	29 (55.8)	48 (29.8)	23 (69.7)	12 (20.7)
Negative	33 (38.8)	159 (72.6)	23 (44.2)	113 (70.2)	10 (30.3)	46 (79.3)
Ki67 expression (%), M (P25, P75)	0.3 (0.2, 0.5)	0.3 (0.2, 0.5)	0.3 (0.2, 0.5)	0.3 (0.2, 0.5)	0.3 (0.2, 0.45)	0.3 (0.2, 0.5)
NACT regimen, (n, %)
Anthracycline-based	4 (4.7)	20 (9.1)	3 (5.8)	15 (9.3)	1 (3.0)	5 (8.6)
Taxane-based	20 (23.5)	20 (9.1)	8 (15.4)	15 (9.3)	12 (36.4)	5 (8.6)
Anthracycline + Taxane	45 (53.0)	170 (77.7)	29 (55.7)	123 (76.4)	16 (48.5)	47 (81.1)
Carboplatin-based	16 (18.8)	9 (4.1)	12 (23.1)	8 (5.0)	4 (12.1)	1 (1.7)
Molecular subtype
HR+/HER2+	24 (28.2)	42 (19.2)	14 (26.9)	37 (23.0)	10 (30.3)	5 (8.6)
HR−/HER2+	28 (33.0)	18 (8.2)	15 (28.9)	11 (6.8)	13 (39.4)	7 (12.1)
HR+/HER2−	24 (28.2)	134 (61.2)	18 (34.6)	96 (59.6)	6 (18.2)	38 (65.5)
HR−/HER2−	9 (10.6)	25 (11.4)	5 (9.6)	17 (10.6)	4 (12.1)	8 (13.8)

cT stage clinical tumor stage, cN stage clinical nodal stage, the maximum tumor diameter at baseline, percentage reduction of tumor maximum diameter from baseline to third cycle of NACT, ER estrogen receptor, PR progesterone receptor, HR hormone receptor, HER2 human epidermal growth factor receptor 2, Ki67 Ki67 proliferative index, NACT neoadjuvant chemotherapy, pCR pathological complete response.

Predictors of pathological complete response (pCR)

Logistic regression was performed to explore predictors of pCR, analyzing potential relations with patient age, , , ER, PR, HER2, Ki67, and NACT regimen (Table 2). In univariate analysis, smaller values were associated with higher probability of pCR (odds ratio ; 95% confidence interval [CI], 0.711–0.965; ). Likewise, HR (ER/PR) status, HER2 status, and NACT regimen (all p < 0.001), as well as Ki67 expression level (), emerged as significant predictors of pCR. To determine the feasibility of predicting therapeutic response through early tumor dynamic changes, we analyzed percentages of change in (as shown by US) relative to baseline values across initial four NACT cycles. In Fig. 2a, values at respective NACT cycles are shown to compare between pCR and non-pCR. Significant differences were apparent in cycles 2–4. Furthermore, was confirmed as a significant predictor of pCR. Despite the importance of early response prediction, the earliest values were not the best. Figure 2b demonstrates that and surpassed and in terms of capacity to predict pCR, which was shown by AUC. Additionally, the ROC shows that and had similar AUC. Due to the aim of early response evaluation, we subsequently selected change in maximum tumor diameter at third cycle as the preferred metric, denoted as (). ROC analysis yielded a cut-off value of 29.9% for .

Table 2.

Univariate and multivariate analyses of pCR.

Characteristic	Univariate analysis			Multivariable analysis
	OR	95% CI	p value	OR	95% CI	p value
Age (year)
> 50	1
≤ 50	1.353	0.819–2.236	0.238
T0	0.828	0.711–0.965	0.016	0.706	0.582–0.857	< 0.001
ΔT3	9.518	3.327–27.232	< 0.001	10.795	3.077–37.880	< 0.001
ER status
Negative	1
Positive	0.259	0.153–0.438	< 0.001	0.417	0.192–0.907	0.027
PR status
Negative	1			1
Positive	0.394	0.235–0.661	< 0.001	0.946	0.440–2.035	0.887
HER2 status
Negative	1			1
Positive	4.176	2.464–7.078	< 0.001	1.641	0.797–3.375	0.179
Ki67 expression	5.573	1.828–16.988	0.003	4.843	1.202–19.519	0.027
NACT regimen			< 0.001
Anthracycline-based	1			1
Taxane-based	5.000	1.448–17.271	0.011	3.083	0.712–13.342	0.132
Anthracycline + Taxane	1.324	0.431–4.067	0.625	1.376	0.391–4.847	0.619
Carboplatin-based	8.889	2.307–34.248	0.001	7.928	1.567–40.112	0.012

OR odds ratio, CI confidence interval, the maximum tumor diameter at baseline, percentage reduction of tumor maximum diameter from baseline to third cycle of NACT, ER estrogen receptor, PR progesterone receptor, HR hormone receptor, HER2 human epidermal growth factor receptor 2, Ki67 Ki67 proliferative index, NACT neoadjuvant chemotherapy, pCR pathological complete response.

p < 0.05 was considered statistically significant.

Fig. 2.

The comparison of , , , and for predicting the probability of pCR. (a) The comparison of between pCR and non-pCR patients at first four cycles of NACT. Significant differences were apparent in cycles 2–4, except for cycle 1. (b) The receiver operating characteristic (ROC) curves of , , , and for predicting the probability of pCR. and had better ability to predict pCR compared to and . To measure the early dynamic changes of the tumor, we defined as . refers to the maximum diameter of the tumor at the baseline. refers to the maximum diameter of the tumor at different cycles of NACT. refers to the cycle of NACT. p < 0.05 was considered statistically significant.

Multivariate logistic regression analysis was next carried out in search of independent predictors for pCR. The forest plot is shown as Fig. 3. Relative to ER+ patients, those with ER- status were more likely to achieve pCR (, ; ). In addition, (, ; ) and () displayed significant associations with pCR, as did Ki67 expression (, ; ). Once adjusted for ER status, Ki67 expression, , and , a significantly higher rate of pCR was also evident after carboplatin-based (vs anthracycline-based) treatment (, ; ). Although significance was not reached, patients receiving taxane-based () or anthracycline plus taxane-based () regimen seemed to have a higher odds ratio than those administered anthracycline-based treatment ().

Fig. 3.

Forest plot of variables for prediction of pCR. ER, estrogen receptor status; Ki67, Ki67 proliferative index; , the maximum tumor diameter at baseline; , percentage reduction of tumor maximum diameter from baseline to third cycle of NACT; NACT regimen, T, Taxanes; A, Anthracycline; Cb, carboplatin; pCR, pathological complete response. p < 0.05 was considered statistically significant.

Nomogram construction and validation

Our nomogram incorporated significant predictors of pCR (, ER status, Ki67 expression, NACT regimen) established by multivariate logistic regression analysis, with the addition of (Fig. 4). Each variable was assigned a point value (top scale), the sum of which reflected pCR probability (bottom scale). Nomogram validation by ROC analysis generated AUC values of 0.75 in both training and validation sets (training set, 95% CI: 0.67–0.83; validation set, 95% CI: 0.64–0.86), suggesting good predictive efficiency (Fig. 5a,b). The calibration curve for pCR probability also showed good agreement between predicted and observed results (Fig. 5c). The Hosmer–Lemeshow test indicated no significant deviation of obtained results from an ideal fitting ().

Fig. 4.

Nomogram for predicting the probability of pCR. Nomogram for predicting the probability of pCR. Variables including NACT regimen, , , ER status, and Ki67 proliferative index. pCR, pathological complete response; NACT regimen, T, Taxanes; A, Anthracycline; Cb, carboplatin; , the maximum tumor diameter at baseline; , percentage reduction of tumor maximum diameter from baseline to third cycle of NACT; ER, estrogen receptor status; Ki67, Ki67 proliferative index.

Fig. 5.

The receiver operating characteristic (ROC) curves and calibration curve of a nomogram. (a) The ROC curves for predicting the probability of pCR in the training set. (b) The ROC curves for predicting the probability of pCR in the validation set. (c) The calibration curve of the nomogram. The dashed line named Ideal presents perfect performance, the other dashed line named Apparent presents the actual performance of the nomogram, and the solid line named Bias-corrected presents the actual performance of the nomogram after bias-correction.

We additionally conducted a DCA to assess predictive performances of four separate models (Fig. 6). Model 1 was developed using ER status, Ki67 expression, and NACT regimen; Model 2 and Model 3 were created using predictors in Model 1 plus or ; and Model 4 encompassed all stated predictors. Ultimately, the greatest net benefit in predicting NACT response was conveyed by Model 4. Relative to Model 2 and Model 3, there were significant gains in predictive efficiency, supporting as a driving factor for improved predictive performance. In Fig. 7, we have presented two patients presented with similar clinical characteristics but with differing tumor dynamic trends during NACT. After three treatment cycles, Patient 2 registered a higher than that of Patient 1 (84% vs 27%). The total points calculated by nomogram of Patient 1 and Patient 2 are respectively 150 and 167, and Patient 2 had a higher probability of pCR compared to Patient 1. Pathological results upon NACT completion were confirmatory of pCR in Patient 2 but not in Patient 1, underscoring the utility as a relatively sensitive index of NACT response.

Fig. 6.

The decision curve analysis (DCA) of four different models. DCA was performed to compare the predictive performance among four different models. Model 1, NACT regimen + ER + Ki67; Model 2, NACT regimen + ER + Ki67 + ; Model 3, NACT regimen + ER + Ki67 + ; Model 4, NACT regimen + ER + Ki67 + +. Model 4 had the highest net benefit to predict NACT response. Compared with Model 2, Model 3 showed a significant improvement in predictive efficiency, indicating that is a driving factor for improving predictive performance.

Fig. 7.

Schematic of dynamic tumor changes during NACT in two patients with similar clinicopathological characteristics (Created by Figdraw). Molecular subtypes were identical, both HR−/HER2 + with high-level Ki67 expression; TNM classifications were the same; and ages were roughly similar (49 vs 53 years). In Patient 2, maximum tumor diameter shrank from 3.2 cm at baseline to 0.5 cm after a third cycle of NACT (Anthracycline, Cyclophosphamide, Trastuzumab, and Pertuzumab [TCHP]), with at 84%. Although unchanged (still 0.5 cm) upon completion of NACT. By contrast, the of Patient 1 was just 27% after a third cycle of NACT (Anthracycline and Cyclophosphamide followed by Taxanes, Trastuzumab, and Pertuzumab [AC-THP]), with maximum tumor diameter remained at 1.0 cm after NACT. The total points calculated by nomogram of Patient 1 and Patient 2 were respectively 150 and 167. Based on the prediction model, Patient 2 had a higher probability of pCR compared to Patient 1. Pathological results upon NACT completion were confirmatory of pCR in Patient 2 but not in Patient 1.

Machine learning model

During NACT, dynamic changes of US images should not be underestimated. We performed several machine learning algorithms to construct the prediction models, including support vector machines (SVM), decision tree, random forest (RF), gradient boosting machine (GBM), and fully connected neural network (Table 3). The results showed that the SVM exhibited the best predictive performance (). Consequently, our final endeavor was constructing a SVM model driven by clinicopathological data and US parameters at each treatment cycle, including maximum tumor diameters and volumes. The model’s precision, recall and F1 scores were 77.0%, 71.0%, and 74.0%, respectively. The confusion matrix is provided as Fig. 8.

Table 3.

The AUC values of different machine learning algorithms for predicting pCR.

Machine learning algorithms	AUC
SVM	0.868
Fully connected neural network	0.827
RF	0.816
GBM	0.805
Decision tree	0.633

SVM support vectors machines, RF random forest, GBM gradient boosting machine.

Fig. 8.

The confusion matrix of the SVM model (a) and the ROC curve of the SVM model (b).

Discussion

Cure is the ultimate goal of treatment in patients with EBC. Breast cancer is a systemic disease requiring systemic intervention, including chemotherapy and targeted therapy³³. However, the therapeutic effects will vary according to individual, some responding to systemic therapy and others left unaffected. Multicenter prospective clinical studies, such as CREATE-X and KATHERINE, have confirmed that intensified follow-up adjuvant therapy in proven non-responders may improve long-term survival^34,35. Early response evaluations during treatment or even at baseline are therefore essential for optimal outcomes.

NACT is an individual drug sensitivity assessment platform. Patients insensitive to standard NACT regimens must be detected as soon as possible to avoid unnecessary toxicity exposures and extra costs. There is an unmet clinical need for a biomarker to distinguish patient subsets by pCR and non-pCR. Non-pCR patients calls for treatment revision or expeditious surgical management.

Imaging methods and liquid biopsy have been previously utilized during a number of studies to evaluate patient response early. In the PHERGain trial, ¹⁸F-fluorodeoxyglucose (FDG)-positron emission tomography (PET) was performed to metabolically detect HER2+ breast cancer patients, enabling dual targeted therapies and avoiding chemotherapy after just two cycles of treatment³⁶. Dynamic contrast-enhanced (DCE)-MRI performed after two cycles of NACT may also help predict pathological responses in the setting of breast cancer³⁷. Liquid biopsies were undertaken in the WSG-ADAPT-TP trial to determine patient responses in early-stage, defined as 30% decline in Ki67 expression or low cellularity³⁸. Also, ctDNA and circulating free DNA (cfDNA) have known merit as predictive and prognostic factors^39–41. The current contention in China is that MRI and PET-computed tomography (CT), as well as secondary biopsies, may be costly and inordinately traumatic for patients.

US imaging is a noninvasive and cost-effective method for examining the breast. In the GeparTrio trial, US served to identify non-responders who initially received taxanes/anthracycline/cyclophosphamide (TAC) regimen. Switching to TAC plus vinorelbine/capecitabine (TAC-NX) improved their prognosis, without cross-drug resistance or NACT prolongation⁴². Similarly, US imaging has facilitated early NACT response prediction in the Neo-ALTTO trial⁴³; and some small-scale studies have focused on its role in early response evaluation^44,45, without a uniform consensus on timing or criteria of such efforts. In the present study, we primarily addressed early tumor shrinkage in the nomogram, validating tumor reduction after three cycles of NACT as a basis for early response determinations. This finding indicated that the concept of early prediction does not imply that the earlier the prediction, the more accurate it will be. The changes of the tumor after three cycles of NACT may prove to be a more accurate predictor of the response than the earlier assessments. Interestingly, the cut-off value of 29.9% is just comparable to the definition of partial response (PR) by RECIST standards (version 1.1)⁴⁶. This outcome readily demonstrates that early US examinations during NACT are beneficial, ensuring timely modifications of therapeutic regimens.

Machine learning has been extensively applied to therapeutic response evaluations, revealing tumor characteristics that escape the naked eye^47–49. One such model has already been reported for handling dynamic longitudinal US imaging data to predict pCR early during NACT in patients with HER2+ breast cancer⁵⁰. Several other studies have concentrated on pre-, mid-, and post-NACT features of US-delineated NACT response, while failing to explore dynamic tumor changes^51,52. Herein, we assessed US parameters at each cycle of NACT to fully capture tumor dynamics in our prediction model. To the best of our knowledge, it is the first time to investigate the predictive efficacy of dynamic changes in ultrasound-measured tumor maximum diameter throughout the entire NACT cycle. We also input static elements, including hormone receptor (ER/PR) status, HER2 status, Ki67 expression, and NACT regimen. By integrating both dynamic and static features, predictive performance was heightened () to levels reached elsewhere, capable of reliably guiding therapeutic decisions. Furthermore, other researchers have reported that MRI has the potential to predict pCR to NACT, with AUC values ranging from 0.7 to 0.9^53,54. This suggests that our US-based prediction models are comparable to other MRI-based prediction models. Although MRI is considered to have the highest sensitivity for breast cancer detection and the high accuracy to predict the response to NACT^13,55, it is a costly procedure, rendering it impractical for frequent scanning during NACT. In comparison, US is a more cost-effective and widely available modality, and therefore more suitable to be the routine NACT monitoring tool.

Clinicopathological parameters also figured prominently in response evaluation. We identified other independent predictors of pCR, including ER status, Ki67 expression, , and NACT regimen. The nomogram reached an AUC of 0.75. Patients with ER− (vs ER+) tumors displayed a comparatively higher rate of pCR, aligning with past studies and perhaps translating to less NACT sensitivity of ER+ patients³². Ki67 is a biomarker of cellular proliferation with predictive ramifications⁵⁶. In our study, the higher the expression level of Ki67, the more likely pCR would be achieved, albeit with no consensus on a related cut-off value. emerged as an independent predictor as well, and larger tumors had worse responses. Carboplatin is a DNA alkylating agent often administered for TNBC, which bears higher rates of BRCA1/2 mutation^57,58. Patients treated with carboplatin and paclitaxel, with or without anthracycline, were more inclined to pCR, as opposed to recipients of anthracycline-based regimens. However, 80% of those treated with carboplatin were HER2+ and additionally received targeted therapy. We did add to the nomogram, and distinctly contributed to improved predictive performance in the subsequent DCA.

There are certain study limitations to acknowledge. The prediction models were constructed based on retrospective data, which may have biases. To eliminate potential biases, patients were enrolled from multiple centers and objective variables were selected for multivariate analysis. This was done in order to minimize the effect of these biases on the results. The incorporation of prospective data from more clinical trials would improve the clinical evidence supporting the validity of our model. Another drawback is the small number of patients enrolled, knowing that various molecular subtypes commonly influence treatment and therapeutic effects. Moreover, because trastuzumab (Herceptin) has been widely used in China since 2017, some patients with HER2+ tumors did not receive targeted therapy. Furthermore, intra- and inter-observer variability is also typically high for US imaging, which may potentially impact the accuracy and reliability of US results. In light of the aforementioned considerations, all radiologists were required to follow standard guidelines and protocols in order to minimize the potential US variability. In the future, more work is planned to refine this prediction model, especially an expanded patient sampling with broader clinical implications. Prospective clinical trials are ultimately needed to optimize our prediction model.

In conclusion, we have proposed a method for timely prediction of NACT responses. This simple graphic representation is intuitive and may assist clinicians in rendering expeditious, individualized therapeutic decisions. We have also constructed an analogous machine learning model that combines dynamic US changes and static clinicopathological characteristics. It may well serve as a reference for preoperative clinical decision-making. Given their overall clinical potential, validation is warranted for both models through future clinical trials.

Author contributions

XW and YZ conceived this study, performed the data analyses and wrote the manuscript. MY and PQ performed the machine learning part. NW, SW, HC, TZ, and YZ assisted revising the manuscript. ZJ, AZ, FY, XW, and FJ provided a portion of the data. JW designed the study and was the director for the fund. All authors read and approved the final manuscript.

Funding

This study was supported by the “1 + X” program cross-disciplinary innovation project (2022JCXKYB07), and Wu Jieping Medical Foundation (320.6750.2022-19-81).

Data availability

The medical data used in this study are not publicly available due to patient privacy considerations. Interested users may request access to the data for research purposes, through contacting the corresponding author. Institutional approvals of data sharing will be required along with signed data use agreements and/or material transfer agreements.

Declarations

Competing interests

The authors declare that they have no competing interests.

Ethics declarations

This is a retrospective study approved by the ethics committee of Second Affiliated Hospital of Dalian Medical University. All methods were performed in accordance with the relevant guidelines and regulations. Informed consent has been obtained from patients.