Development and validation of machine learning models for predicting HER2-zero and HER2-low breast cancers

Abstract

Objectives

To develop and validate machine learning models for human epidermal growth factor receptor 2 (HER2)-zero and HER2-low using MRI features pre–neoadjuvant therapy (NAT).

Methods

Five hundred and sixteen breast cancer patients post-NAT surgery were randomly divided into training (n = 362) and internal validation sets (n = 154) for model building and evaluation. MRI features (tumour diameter, enhancement type, background parenchymal enhancement, enhancement pattern, percentage of enhancement, signal enhancement ratio, breast oedema, and apparent diffusion coefficient) were reviewed. Logistic regression (LR), support vector machine (SVM), k-nearest neighbour (KNN), and extreme gradient boosting (XGBoost) models utilized MRI characteristics for HER2 status assessment in training and validation datasets. The best-performing model generated a HER2 score, which was subsequently correlated with pathological complete response (pCR) and disease-free survival (DFS).

Results

The XGBoost model outperformed LR, SVM, and KNN, achieving an area under the receiver operating characteristic curve (AUC) of 0.783 (95% CI, 0.733-0.833) and 0.787 (95% CI, 0.709-0.865) in the validation dataset. Its HER2 score for predicting pCR had an AUC of 0.708 in the training datasets and 0.695 in the validation dataset. Additionally, the low HER2 score was significantly associated with shorter DFS in the validation dataset (hazard ratio: 2.748, 95% CI, 1.016-7.432, P = .037).

Conclusions

The XGBoost model could help distinguish HER2-zero and HER2-low breast cancers and has the potential to predict pCR and prognosis in breast cancer patients undergoing NAT.

Advances in knowledge

HER2-low–expressing breast cancer can benefit from the HER2-targeted therapy. Prediction of HER2-low expression is crucial for appropriate management. MRI features offer a solution to this clinical issue.

Introduction

Pathological complete response (pCR) after neoadjuvant therapy (NAT) can be used as a surrogate endpoint of survival in human epidermal growth factor receptor 2 (HER2)–positive breast cancer.¹ Targeted therapies specifically inhibiting the HER2 pathway improved pCR rates and survival outcomes in patients with HER2-positive (immunohistochemistry [IHC] score 3+ or IHC2+/in situ hybridization [ISH] positive) but were not applied in traditionally HER2-negative patients (IHC 0, 1+, or 2+/ISH negative).^2–5 Recent clinical trials have shown that the novel anti-HER2–targeting antibody-drug conjugates (ADCs) can benefit subgroups of breast cancers with low HER2 expression (IHC1+ or IHC2+/ISH negative), which are usually classified as HER2-negative.^6–8 Recent clinical investigations have indicated that HER2-low–expression breast cancers present with distinctive clinical behaviour, divergent rates of pCR, and varied prognostic outcomes when compared with HER2-zero expression.^9,10

The expression of HER2 leads to abnormal activation of downstream signalling pathways, resulting in uncontrolled and unrestricted cell growth.¹¹ It is plausible to hypothesize that increased cell proliferation may manifest as more diverse spatial relationships and voxel intensity characteristics in imaging phenotypes. MRI features have been used to assess the intratumoural heterogeneity of breast cancer, enabling correlation with tumour subtypes, NAT response, and prognosis.^12–14 While several studies have demonstrated that MRI radiomic signatures have the potential to distinguish HER2 expression,^15–17 these radiomics studies usually require complicated image processing and analysis, posing challenges for clinical use. Simple MRI descriptors have been validated in effectively differentiating HER2-positive and HER2-low breast cancers.¹⁸ However, these relatively accessible MRI features have not been sufficiently analysed to exploit the potential of evaluating HER-zero and HER-low.

Machine learning (ML) has found increased applications in biomedicine, such as biological analysis, image processing, and classification.¹⁹ ML methods have proven particularly effective in key feature detection and model construction.¹⁹ Therefore, this study aimed to develop and validate ML models that utilize routinely collected pretreatment MRI characteristics to assess the HER2 status in breast cancers before NAT.

Methods

Patient selection and data collection

This study was reviewed and approved by the local institutional review board, which waived the requirement for written informed consent due to the analysis’ retrospective nature. Medical records of breast cancer patients treated from 1 April 2015 to 30 June 2021 were retrieved from clinical databases for analysis. Patients with pathologically proven invasive breast cancer who underwent surgery after NAT were included in the study. Exclusion criteria were HER2-positive cancer or unknown HER2 status, unavailability of images from MRI before NAT, poor MRI quality, recurrent breast cancer in women with a history of breast cancer, and/or multiple primary neoplasms, distant metastasis at the initial diagnosis, and bilateral breast cancer. Finally, 516 patients were included for further analysis. The flowchart of the study is shown in Figure 1.

Figure 1.

Flowchart of the study population. Abbreviations: HER2 = human epidermal growth factor receptor 2; NAT=neoadjuvant therapy.

Clinical and pathological characteristics were collected for analysis, including age, menopausal status, histologic type, histologic grade, HER2 status, Ki-67 index, hormone receptor (HR), NAT regiment, breast surgery, and adjuvant therapy. The tumour type and receptor status were confirmed through IHC and ISH analysis of core biopsies guided by ultrasound. HER2 status was evaluated following the 2018 ASCO/CAP guidelines.³ Tumours with an IHC 0 score were classified as HER2-zero. Tumours with an IHC score of 1+ or 2+/ISH negative were classified as HER2-low.^4,20 Ki-67 expression was evaluated as either low or high, with <30% considered low expression and ≥30% considered high expression.²¹ Hormone receptor was defined as positive if tumour cells exhibited ≥1% of nuclear staining of oestrogen receptor or progesterone receptor.²²

Pathological complete response was defined as a complete absence of invasive tumour cells both in the breast and axillary lymph nodes (ypT0/N0 or ypTis/N0),¹ based on the pathological examination of the surgical specimen. Disease-free survival (DFS) was defined as the time in months between the date of surgery and the occurrence of events including invasive local, regional, and distant recurrences; contralateral breast cancer; second nonbreast primary cancer; death from any cause; and in situ cancer events.²³ The cut-off date for follow-up was June 30, 2023.

MRI acquisition and analysis

MRI was conducted using a 1.5-T GE Optima MR360 MRI scanner (GE Healthcare) and a 1.5-T Philips Achieva MRI scanner (Philips). The detailed parameters of breast MRI acquisition are shown in Table S1. Two board-certified radiologists with 7 years of interpreting breast MRI studies, blinded to the histopathologic results, reviewed the MRI independently. To resolve discordant image interpretations, a senior radiologist with 18 years of expertise in breast imaging was consulted to facilitate a consensus through discussion. Brest MRI were evaluated according to the 2013 Breast Imaging-Reporting and Data System (BI-RADS) MRI lexicon.²⁴ The morphologic and kinetic features of findings were described using the BI-RADS lexicon. MRI features of the breast tumour were assessed in terms of tumour diameter, lesion enhancement type (mass, nonmass enhancement), background parenchymal enhancement, enhancement pattern (homogenous, heterogeneous, and rim enhancement), percentage of enhancement (PE),²⁵ signal enhancement ratio (SER),²⁵ breast oedema,²⁶ and apparent diffusion coefficient (ADC) value.

Tumour diameter was defined as the maximum diameter at the largest cross section of a transverse contrast-enhanced T1-weighted MRI scan. On dynamic contrast-enhanced (DCE) imaging, lesions were categorized as exhibiting either mass or nonmass enhancement, and further distinguished based on their homogeneity or heterogeneity, as well as the presence or absence of rim enhancement. These classifications were made in accordance with the guidelines provided by the 2013 BI-RADS lexicon. Background parenchymal enhancement was classified as minimal, mild and moderate, or marked. For analysis of lesion enhancement kinetics, 2D regions of interest (ROI) were manually drawn for the entire lesion on the largest cross-sectional slice, and areas of necrosis and cystic were avoided as much as possible. The mean value was computed by aggregating the ROIs across all voxels within the lesion. Tumour contrast enhancement shown on the MRI was evaluated quantitatively by several parameters as previously described²⁵: (1) PE and (2) SER. PE shows the degree of signal enhancement within the tumour at 120 s after contrast delivery, calculated by PE=(S1−S0)/S0. Where S0 is the baseline signal intensity before contrast administration and S1 is the signal intensity measured at 120 s after contrast delivery. SER is a unitless index that reflects the rate of contrast washout in the tumour between 120 and 300 s after contrast delivery, calculated by: SER=(S1−S0)/(S2−S0). Where S2 is the MRI signal intensity at 300 s after contrast delivery. Breast oedema was divided into 4 grades based on T2WI: no oedema, peritumoural oedema, prepectoral oedema, and subcutaneous oedema.^26,27 The mean and minimum ADC values (ADC_mean and ADC_min) of the tumour area were measured, corresponding to the lesion identified on DCE imaging.

Model construction, validation, and statistical analysis

Patients were randomly divided into training and internal validation sets for model building and evaluation. Models with 10-fold cross-validation were constructed including logistic regression (LR), support vector machine (SVM) classifier, k-nearest neighbour (KNN), and extreme gradient boosting (XGBoost). The discriminative ability of all developed models was initially evaluated on the internal validation set. The overall performance of each model was assessed using the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value. CIs were generated, and pairwise comparisons of the AUC between models were conducted using the method described by DeLong et al.²⁸ The Hosmer-Lemeshow goodness-of-fit test was used to test model calibration. The Shapley additive explanation method was employed to elucidate the significance of features in the ML model.²⁹ The HER2 score was calculated based on the best-performing model among the 4 models and utilized to predict pCR. The optimal threshold of HER2 score, calculated by the maximally selected rank statistics, divided the patients into 2 groups: high-risk DFS and low-risk DFS.

All statistical analyses were performed using R software (version 4.1.3; https://www.r-project.org/) and IBM SPSS Statistics 23 (IBM Corporation, Armonk, NY, USA). Baseline clinicopathological characteristics were compared with a Wilcoxon rank-sum test for continuous variables, and a chi-square test, or Fisher’s test for categorical variables. All variables associated with HER2 expression at the P < .05 level in univariable analysis were included in multivariable LR analysis. We evaluated the interobserver agreement for subjective MRI characteristics using intraclass correlation coefficient (ICC). Two-sided P < .05 was considered statistical significance.

Results

Clinicopathological and MRI characteristics of study patients

A total of 516 breast cancer patients met the inclusion criteria and were included in the analysis. The HER2-low status was verified in 68.8% (249 out of 362) of patients in the training dataset and 64.9% (100 out of 154) of patients in the validation dataset. The interobserver agreements for the MRI characteristics ranged between an ICC of 0.523 and 0.920. The clinicopathological and MRI characteristics of both the training dataset and the validation dataset are summarized in Table 1. Compared with the training set, patients from the internal validation dataset had a high rate of no special type of invasive carcinoma (98.7% [152 of 154] vs 94.8% [343 of 362]; P = .038). No significant difference was observed in other clinicopathological and MRI findings between the training and validation datasets (Table 1).

Table 1.

Clinicopathological and MRI characteristics of all patients.

Characteristic	Total patients (n = 516)	Training dataset (n = 362)	Validation dataset (n = 154)	P-value
Age (years)	49 (23-87)	49 (24-77)	48 (23-87)	.752
Menopausal status				.958
Premenopause	304 (58.9%)	213 (58.8%)	91 (59.1%)
Postmenopause	212 (41.1%)	149 (41.2%)	63 (40.9%)
Location				.274
L	277 (53.7%)	200 (55.2%)	77 (50.0%)
R	239 (46.3%)	162 (44.8%)	77 (50.0%)
Clinical T stage				.247
T1-T2	388 (75.2)	267 (73.8%)	121 (78.6%)
T3-T4	128 (24.8)	95 (26.2%)	33 (21.4%)
Clinical N stage				.690
N0	130 (25.2%)	93 (25.7%)	37 (24.0%)
N1-3	386 (74.8%)	269 (74.3%)	117 (76.0%)
Histologic type				.038
Invasive carcinoma of no special type	495 (95.9%)	343 (94.8%)	152 (98.7%)
Other invasive histology	21 (4.1%)	19 (5.2%)	2 (1.3%)
Histologic grade				.529
1 or 2	289 (56.0%)	206 (56.9%)	83 (53.9%)
3	227 (44%)	156 (43.1%)	71 (46.1%)
HER2 status				.392
Zero	167 (32.4%)	113 (31.2%)	54 (35.1%)
Low	349 (67.6%)	249 (68.8%)	100 (64.9%)
HR status				.332
Negative	132 (25.6%)	97 (26.8%)	35 (22.7%)
Positive	384 (74.4%)	265 (73.2%)	119 (77.3%)
Ki-67 expression				.483
Low (<30%)	145 (28.1%)	105 (29.0%)	40 (26.0%)
High (≥30%)	371 (71.9%)	257 (71.0%)	114 (74.0%)
NAT regiment				.175
Anthracycline-based	20 (3.9%)	12 (3.3%)	8 (5.2%)
Taxane-based	149 (28.9%)	102 (28.2%)	47 (30.5%)
Anthracycline and Taxane-based	297 (57.6%)	215 (59.4%)	82 (53.2%)
Other chemotherapy	10 (1.9%)	4 (1.1%)	6 (3.9%)
Endocrine therapy	40 (7.8%)	29 (8.0%)	11 (7.1%)
Breast surgery				.597
Conserving surgery	113 (21.9%)	77 (21.3%)	36 (23.4%)
Total mastectomy	403 (78.1%)	285 (78.7%)	118 (76.6%)
Adjuvant chemotherapy				.153
Yes	276 (53.5%)	188 (51.9%)	88 (57.1%)
No	227 (44.0%)	162 (44.8%)	65 (42.2%)
Missing	13 (2.5%)	12 (3.3%)	1 (0.6%)
Adjuvant radiotherapy				.966
Yes	324 (62.8%)	226 (62.4%)	98 (63.6%)
No	158 (30.6%)	112 (30.9%)	46 (29.9%)
Missing	34 (6.6%)	24 (6.6%)	10 (6.5%)
pCR				.536
Negative	401 (77.7%)	284 (78.5%)	117 (76.0%)
Positive	115 (22.3%)	78 (21.5%)	37 (24.0%)
Enhancement type				.739
Mass	420 (81.4%)	296 (81.8%)	124 (80.5%)
Nonmass	96 (18.6%)	66 (18.2%)	30 (19.5%)
BPE				.498
Minimum or mild	392 (76.0%)	272 (75.1%)	120 (77.9%)
Moderate or maximum	124 (24.0%)	90 (24.9%)	34 (22.1%)
Enhancement pattern				.990
Homogeneous	29 (5.6%)	20 (5.5%)	9 (5.8%)
Heterogeneous	450 (87.2%)	316 (87.3%)	134 (87.0%)
Rim enhancement	37 (7.2%)	26 (7.2%)	11 (7.1%)
Oedema				.470
No oedema	122 (23.6%)	91 (25.1%)	31 (20.1%)
Peritumoral oedema	122 (23.6%)	82 (22.7%)	40 (26.0%)
Prepectoral oedema	139 (26.9%)	100 (27.6%)	39 (25.3%)
Subcutaneous oedema	133 (25.8%)	89 (24.6%)	44 (28.6%)
ADCmin (10−3mm2/s)	0.57 (0.14-1.53)	0.57 (0.14-1.53)	0.57 (0.25- 1.16)	.482
ADCmean (10−3mm2/s)	0.91 (0.43-1.75)	0.92 (0.43-1.75)	0.90 (0.54-1.64)	.505
Tumour diameter (mm)	30 (9-121)	30 (9-121)	28 (9-75)	.104
PE	1.62 (0.46-3.14)	1.64 (0.46-3.13)	1.59 (0.55-3.14)	.196
SER	1.04 (0.56-1.37)	1.04 (0.56-1.37)	1.05 (0.67-1.34)	.971

Data were presented as n (%) or median (range).

Abbreviations: ADC_min = the minimum value of apparent diffusion coefficient; ADC_mean = the mean value of apparent diffusion coefficient; BPE = background parenchymal enhancement; HR = hormone receptor; HER2 = human epidermal growth factor receptor; NAT = neoadjuvant therapy; pCR = pathological complete response; PE = percentage of enhancement; SER = signal enhancement ratio.

The contrast between patients with HER2-zero and HER2-low status in the training dataset is elucidated in Table 2. Regarding clinicopathological factors, differences were observed between the HER2-zero group and the HER2-low group for the following variables: (1) Histologic grade (histologic grade 3 in HER2-zero vs HER2-low: 56.6% vs 36.9%; P < .001); (2) HR-positive (HER2-zero vs HER2-low: 57.5% vs 80.3%; P < .001); (3) high ki-67 expression (HER2-zero vs HER2-low: 80.5% vs 66.7%; P = .007). As for MRI characteristics, ADC_mean, PE, and SER were significantly different between HER2-zero and HER2-low groups. None of the remaining clinicopathological and MRI characteristics were different between the HER2-zero and HER2-low groups in the training dataset. In multivariable analysis, after adjustment for the variables that presented a significantly uneven distribution in clinicopathological and MRI characteristics (histologic grade, Ki-67 status, PE), HR status, ADC_mean, and SER remained statistically significant (Table S2). Table S3 shows the relationship between pCR and clinicopathological characteristics of all patients. The pCR group exhibited a lower clinical stage, higher histologic grade, more HER2-zero status, more HR negative, and higher Ki-67 expression than the non-pCR group. Patients who achieved pCR presented with more taxane-based and less endocrine therapy.

Table 2.

Clinicopathological and MRI characteristics of HER2-zero and HER2-low patients in the Training Datasets.

Characteristic	HER-zero (n = 113)	HER-low (n = 249)	P-value
Age (years)	48 (32-71)	49 (24-77)	.199
Menopausal status			.418
Premenopause	70 (61.9%)	143 (57.4%)
Postmenopause	43 (38.1%)	106 (42.6%)
Location			.215
L	57 (50.4%)	143 (57.4%)
R	56 (49.6%)	106 (42.6%)
Clinical T stage			.145
T1-T2	89 (78.8%)	178 (71.5%)
T3-T4	24 (21.2%)	71 (28.5%)
Clinical N stage			.789
N0	28 (24.8%)	65 (26.1%)
N1-3	85 (75.2%)	184 (73.9%)
Histologic type			.326
Invasive carcinoma of no special type	109 (96.5%)	234 (94.0%)
Other invasive histology	4 (3.5%)	15 (6.0%)
Histologic grade			<.001
1 or 2	49 (43.4%)	157 (63.1%)
3	64 (56.6%)	92 (36.9%)
HR status			<.001
Negative	48 (42.5%)	49 (19.7%)
Positive	65 (57.5%)	200 (80.3%)
Ki-67 expression			.007
Low (< 30%)	22 (19.5%)	83 (33.3%)
High (≥ 30%)	91 (80.5%)	166 (66.7%)
NAT regiment			.063
Anthracycline-based	5 (4.4%)	7 (2.8%)
Taxane-based	35 (31.0%)	67 (26.9%)
Anthracycline and Taxane-based	68 (60.2%)	147 (59.0%)
Other chemotherapy	2 (1.8%)	2 (0.8%)
Endocrine therapy	3 (2.7%)	26 (10.4%)
Breast surgery			.411
Conserving surgery	27 (23.9%)	50 (20.1%)
Total mastectomy	86 (76.1%)	199 (79.9%)
Adjuvant chemotherapy			.872
Yes	57 (50.4%)	131 (52.6%)
No	52 (46.0%)	110 (44.2%)
Missing	4 (3.5%)	8 (3.2%)
Adjuvant radiotherapy			.232
Yes	77 (68.1%)	149 (59.8%)
No	28 (24.8%)	84 (33.7%)
Missing	8 (7.1%)	16 (6.4%)
pCR			.066
Negative	82 (72.6%)	202 (81.1%)
Positive	31 (27.4%)	47 (18.9%)
Enhancement type			.638
Mass	94 (83.2%)	202 (81.1%)
Nonmass	19 (16.8%)	47 (18.9%)
BPE			.812
Minimum or mild	84 (74.3%)	188 (75.5%)
Moderate or maximum	29 (25.7%)	61 (24.5%)
Enhancement pattern			.778
Homogeneous	5 (4.4%)	15 (6.0%)
Heterogeneous	99 (87.6%)	217 (87.1%)
Rim enhancement	9 (8.0%)	17 (6.8%)
Oedema			.641
No oedema	24 (21.2%)	67 (26.9%)
Peritumoral oedema	29 (25.7%)	53 (21.3%)
Prepectoral oedema	32 (28.3%)	68 (27.3%)
Subcutaneous oedema	28 (24.8%)	61(24.5%)
ADCmin (10−3 mm2/s)	0.55 (0.20-0.96)	0.59 (0.14-1.53)	.066
ADCmean (10−3 mm2/s)	0.87 (0.51-1.41)	0.94 (0.43-1.75)	.009
Tumour diameter (mm)	31 (9-114)	30 (10-121)	.777
PE	1.55 (0.46-3.08)	1.67 (0.50-3.13)	.043
SER	1.01 (0.69-1.27)	1.05 (0.56-1.37)	.003

Data were presented as n (%) or median (range).

Abbreviations: ADC_min = the minimum value of apparent diffusion coefficient; ADC_mean = the mean value of apparent diffusion coefficient; BPE = background parenchymal enhancement; HR = hormone receptor; HER2 = human epidermal growth factor receptor; NAT = neoadjuvant therapy; pCR = pathological complete response; PE = percentage of enhancement; SER = signal enhancement ratio.

Predictive performance of ML algorithms

Utilizing patient characteristics collected before receiving NAT, supervised ML models were trained to identify patients with low levels of HER2 expression. Table 3 provides a summary of the predictive performance of all 4 evaluated classification algorithms in both datasets. Among these algorithms, the XGBoost algorithm exhibited the highest performance for assessing HER2 status, with AUC of 0.783 (95% CI, 0.733-0.833) and 0.787 (95% CI, 0.709-0.865) in the training and validation datasets, which surpass LR (0.622, 95% CI, 0.562-0.682; 0.536, 95% CI, 0.442-0.630), SVM classifier (0.745, 95% CI, 0.691-0.800; 0.776, 95% CI, 0.699-0.853), and KNN models (0.738, 95% CI, 0.688-0.788; 0.756, 95% CI, 0.682-0.830; Figure 2). Comparing the AUC of the XGBoost model with that of the LR model revealed a statistically significant difference in the training and validation datasets. Meanwhile, the AUCs of the SVM and KNN models exhibited no statistically significant disparity in the training and validation datasets. Figure 3 illustrates the ranking of feature importance for the imputed variables in the XGBoost model. The ADC_mean, PE, and SER were the important MRI features (Figure 4). After incorporating the clinicopathological variables (HR status, histologic grade, and Ki-67 status), the XGBoost model attained AUC values of 0.835 and 0.827 in the training and validation datasets, respectively (Figure S1). Finally, the predicted probabilities of the XGBoost model closely aligned with the observed probabilities and showed a good calibration (Hosmer-Lemeshow test: P-value in the training and validation cohorts were .108 and .168 respectively; Figure S2).

Table 3.

Performance of machine learning models in the training and validation datasets.

	The training dataset				The validation dataset
	LR	SVM	KNN	XGBoost	LR	SVM	KNN	XGBoost
AUC	0.622 (0.562-0.682)	0.745 (0.691-0.800)	0.738 (0.688-0.788)	0.783 (0.733-0.833)	0.536 (0.442-0.630)	0.776 (0.699-0.853)	0.756 (0.682-0.830)	0.787 (0.709-0.865)
Accuracy, %	54.7	71.8	66.6	71.0	47.4	73.4	67.5	72.1
Sensitivity, %	45.0	69.9	62.2	70.3	26.0	68.0	61.0	68.0
Specificity, %	76.1	76.1	76.1	72.6	87.0	83.3	79.6	79.6
PPV, %	80.6	86.6	85.2	85.0	78.8	88.3	84.7	86.1
NPV, %	38.6	53.4	47.8	52.6	38.8	58.4	52.4	57.3
P-value	<.0001	.219	.104	Ref	.0006	.836	.514	Ref

Abbreviations: AUC = area under the curve; LR = logistic regression; KNN = k-nearest neighbour; NPV = negative predictive value; PPV = positive predictive value; ref = reference; SVM = support vector machine; XGBoost = extreme gradient boosting.

Figure 2.

Area under the receiver operating characteristic curves are shown for all 4 models. (A) The training dataset; (B) the validation dataset. Abbreviations: AUC = area under the receiver operating characteristic curve; KNN = k-nearest neighbour; LR = logistic regression; XGB = extreme gradient boosting

Figure 3.

Feature importance plot for the XGBoot model in the training dataset. The yellow and purple in each row indicate participants with low to high values of the particular variable, while the X-axis gives the SHAP values for the impact on the model. Abbreviations: SHAP = Shapley additive explanation values; PE = percentage of enhancement; SER = signal enhancement ratio.

Figure 4.

Percentage of enhancement (PE), signal enhancement ratio (SER), and the mean value of apparent diffusion coefficient (ADC_mean) in human epidermal growth factor receptor 2 (HER2)-low and HER2-zero patients. (A-D) HER2-low patient, PE = 1.92; SER = 1.06; ADC_mean = 1.002. (E-H) HER2-zero patient, PE = 0.47; SER = 0.72; ADC_mean = 0.83. (A, E) S0, the baseline signal intensity; (B, F) S1, the signal intensity measured at 120 s after contrast delivery; (C, G) S2, the signal intensity measured at 300 s after contrast delivery; (D, H): ADC. PE = (S1−S0)/S0; SER=(S1−S0)/(S2−S0).

Pathological complete response and disease-free survival analysis

The HER2 score established by the XGBoost model, which was used for predicting pCR, achieved an AUC of 0.708 in the training datasets and an AUC of 0.695 in the validation dataset (Figure 5A and B). In exploratory survival analysis, with a median follow-up of 40 months (median [IQR], 39.9 [33.8-46.0] months), 70 patients had recurrence events in the whole dataset. The HER2 score was further employed to make predictions regarding DFS in a cohort of breast cancer patients undergoing neoadjuvant chemotherapy. The Kaplan-Meier survival curves demonstrated that the low HER2 score was significantly associated with a shorter DFS in the validation dataset (HR: 2.748, 95% CI: 1.016-7.432, P = .037; Figure 6B). However, no significant difference was found in the training set (HR: 1.561, 0.811-3.003, P = .18; Figure 6A).

Figure 5.

Area under the receiver operating characteristic curve of pathological complete response. (A) The training dataset; (B) the validation dataset. Abbreviation: AUC = area under the receiver operating characteristic curve.

Figure 6.

Kaplan-Meier curve of disease-free survival in a cohort of breast cancer patients undergoing neoadjuvant chemotherapy. (A) The training dataset; (B) the validation dataset. Abbreviation: HER2 = human epidermal growth factor receptor 2.

Discussion

This study investigated the performance of an MRI-based model in differentiating between HER2-low and HER2-zero breast cancers. The XGBoost model obtained the best classification in both the training and internal validation cohorts. The MRI features derived from the tumour exhibited the capacity to distinguish between HER2-low and HER2-zero breast cancers (internal validation cohorts AUC, 0.787). Our results revealed that the MRI characteristics, including PE, ADC_mean, and SER, significantly contribute to the differentiation between HER2-zero and HER2-low patients. Furthermore, the model demonstrated predictive capabilities for both pCR status and DFS.

The HER2-low expression has now been recognized as targetable with the novel anti-HER2 ADCs, which have been shown to improve progression-free survival and overall survival in the phase III DESTINYBreast04 trial.⁶ Previous studies used MRI to detect and characterize associations between imaging phenotypes and HER2 expression by DCE-MRI.^30–33 In our study, we observed that the kinetic characteristics, namely the PE and SER obtained from DCE-MRI were higher in the HER2-low breast cancer cohort compared to the HER2-zero. This observation suggests a plausible association between HER2 expression and the up-regulation of vascular endothelial growth factor in human breast cancer cells.^34,35 Additionally, we noted that the HER2-low group demonstrated a higher ADC value compared to the HER2-zero group. The correlation between HER2 expression and increased angiogenesis potentially led to enhanced perfusion, resulting in elevated ADC values.^36,37 Furthermore, our findings showed the HER2-low cohort was more likely to be HR-positive, had lower histologic grade, and exhibited lower Ki-67 expression, which aligns with the findings of several other studies.^9,38,39

In this study, the XGBoost model, built using MRI characteristics, demonstrated superior performance in HER2 status assessment compared to other models such as KNN, SVM, and LR. As a gradient-boosted tree-based ML algorithm, XGBoost has consistently showcased exceptional performance across diverse clinical tasks, surpassing conventional algorithms like random forest, LR, and SVM.⁴⁰ Notably, the XGBoost model employs a gain value to quantify the extent to which a variable enhances the accuracy of predictions, with higher gain values indicating greater improvement in prediction generation. Farrokhian et al also employed the XGBoost model to attain superior performance than conventional algorithms in clinical research.⁴¹ In our study, PE emerged as the most influential factor in the XGBoost model. SER derived from DCE-MRI and ADC value played a crucial role in the predictive model, enabling effective differentiation between these 2 subgroups. A recent study found that the combination of selected features from T1-DCE and ADC achieved the best performance in discriminating HER2 status.¹⁵ Furthermore, we observed that the incorporation of clinicopathological variables contributed to enhancing the discriminative performance of the model (internal validation cohorts AUC, 0.827).

In our research, the routinely collected pretreatment MRI characteristics identified through ML algorithms prove effective in differentiating between HER2-zero and HER2-low breast cancer. While radiomics studies may offer valuable insights for HER2 status assessment,^15,16 their complex analytical process often limits their practical application in clinical practice. The objective MRI characteristics employed in our model are easily obtained and have the potential to streamline clinical workflows. Furthermore, the HER2 score derived from the XGBoost model, utilized for predicting pCR, carries substantial implications for advancing patient management. Moreover, this HER2 score, established via the XGBoost model, exhibited an association with DFS in both the training and validation datasets of patients undergoing neoadjuvant chemotherapy. In the validation dataset, lower HER2 scores tend to have a worse DFS, which is consistent with previous research.^9,38 This contrast in DFS based on the HER2 score was not observed in the training dataset. Considering the prior study’s reliance on large sample sizes, future investigations necessitate an expanded sample size. This step is crucial to substantiate whether the model genuinely predicts distinct prognoses between HER2-low-positive and HER2-zero patients.

Limitations

In our study, several limitations were observed. Primarily, it was a retrospective single-institution analysis, posing challenges in controlling all variables to prevent biases. Additional multicentre investigations involving larger cohorts are imperative to validate the discriminative efficacy. Secondly, the use of a 2D region of interest during the assessment of kinetic features and ADC value was limited, as a 3D volume of interest would provide more comprehensive information.

Conclusion

In conclusion, the preoperative breast MRI characteristics using the XGBoost model could aid in identifying suitable candidates for HER2-targeted therapy. Furthermore, this model shows the potential to predict pCR status and prognosis in breast cancer patients treated with NAT.

Author contributions

Drs X. Huang, L. Wu, and Y. Liu contributed equally and shared the first authorship. Drs. Z. Liu and C. Liang contributed equally to this work as co-corresponding authors.

Supplementary material

Supplementary material is available at BJR online.

Funding

This work was funded by the Key-Area Research and Development Program of Guangdong Province (No. 2021B0101420006); National Natural Science Foundation of China (No. 82071892, 82271941, 82272088); Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application (No. 2022B1212010011); the National Science Foundation for Young Scientists of China (No. 82102019); Medical Scientific Research Foundation of Guangdong Province (No. A2024403).

Conflicts of interest

The authors declare that they have no conflict of interest.