Development and validation of novel machine learning-based prognostic models and propensity score matching for comparison of surgical approaches in mucinous breast cancer: a multicenter study

doi:10.21203/rs.3.rs-4996937/v1

Background Mucinous breast cancer (MBC) is a rare subtype of breast cancer with specific clinicopathologic and molecular features. Despite MBC patients generally having a favorable survival prognosis, there is a notable absence of clinically accurate predictive models.

Methods 7553 patients diagnosed with MBC from the SEER database spanning 2010 to 2020 were included for analysis. Cox regression analysis was conducted to identify independent prognostic factors. Ten machine learning algorithms were utilized to develop prognostic models, which were further validated using MBC patients from two Chinese hospitals. Cox analysis and propensity score matching were applied to evaluate survival differences between MBC patients undergoing mastectomy and breast-conserving surgery (BCS).

Results We determined that the XGBoost models were the optimal models for predicting overall survival (OS) and breast cancer-specific survival (BCSS) in MBC patients with the most accurate performance (AUC = 0.833–0.948). Moreover, the XBGoost models still demonstrated robust performance in the external test set (AUC = 0.856–0.911). We also developed an interactive web application to facilitate the utilization of our models by clinicians or researchers. Patients treated with BCS exhibited superior OS compared to those undergoing mastectomy (p < 0.001, HR: 0.60, 95% CI: 0.47–0.77). However, no significant difference was observed in the risk of breast cancer-related mortality. Furthermore, we identified a significant improvement in OS for patients aged 66 or older, white, divorced, with a household income exceeding $40,000, of grade I, HR+/HER2-, with T1 and T2 tumors, and not receiving chemotherapy when treated with BCS.

Conclusion We have successfully developed 6 optimal prognostic models utilizing the XGBoost algorithm to accurately predict the survival of MBC patients. The external validation confirmed the high generalizability of our models. Notably, we observed a significant improvement in OS for patients undergoing BCS.

Biological sciences/Cancer

Health sciences/Medical research

Health sciences/Oncology

mucinous breast cancer

machine learning

prognosis

surgery

propensity score matching

Mucinous breast cancer (MBC) is a rare pathological subtype of breast cancer (BC), accounting for approximately 2–5% of all BC cases^[1]. Despite its rarity, the prevalence of breast cancer globally has been steadily increasing, resulting in a rapid rise in the number of individuals diagnosed with MBC^{[2, 3]}. Compared with other ordinary types of BC such as infiltrating ductal carcinoma (IDC), MBC exhibits specific clinicopathologic and molecular features. These include a higher proportion of hormone receptor expression and a lower rate of lymph node metastasis^[4–8]. MBC is more commonly diagnosed in postmenopausal women and generally exhibits a favorable survival outcome^{[9, 10]}. Due to limited clinical data, the systemic treatment of MBC is mostly guided by the experience gained from treating IDC^{[11, 12]}.

Previous studies have built a few nomograms to predict prognosis in early MBC^[13–15]. Due to the low incidence of MBC, prognostic models have been constructed based on the single Surveillance Epidemiology and End Results (SEER) database, without external dataset to validate the extensiveness of the models. Additionally, the performance of these models was not satisfactory (area under the curve (AUC) value or C-index is about 0.7–0.8). Machine learning (ML) is an emerging field in medicine that encompasses a powerful set of algorithms capable of representing, adapting, learning, predicting, and analyzing data^{[16, 17]}. Extreme gradient boosting (XGBoost) is a variant of the gradient boosting tree algorithm, which updates the model by iteratively calculating the negative gradient of the loss function in a stepwise iterative manner, making the model's prediction results gradually approach the true value^[18]. XGBoost has gradually gained popularity in the medical field for disease prediction, diagnostic assistance, and medical risk assessment. However, XGBoost has not yet been applied in the prediction of MBC prognosis.

The treatment of MBC patients lacks sufficient evidence and standardized guidelines. Currently, mastectomy and breast-conserving surgery (BCS) are the mainstream surgical approaches for treating MBC. An observational study has concluded that patients undergoing BCS have a better prognosis than those undergoing mastectomy^[19]. However, due to the inability to perform randomized allocation, retrospective observational studies are invariably affected by selection bias, which can affect the reliability of the results. Propensity score matching (PSM) is commonly utilized to achieve matching between study and control groups, thereby minimizing potential between-group differences. The survival benefit of different surgical approaches for MBC patients has not been confirmed after PSM.

In this study, we developed models to predict overall survival (OS) and breast cancer-specific survival (BCSS) of MBC patients based on ten ML algorithms using the SEER database. We also retrospectively collected MBC patients data from two Chinese hospitals to assess the applicability of the models. In addition, PSM was used to investigate the survival benefit between patients who underwent mastectomy and those who underwent BCS. The findings of this study aim to provide prognostic assessment and personalized treatment options for MBC patients through the optimal model for predicting prognosis.

Patients and study design

The design of our study has been depicted using the flowchart (Fig. 1). This study collected patient information from three sources. The SEER database is a public database created and maintained by the National Cancer Institute. First,this study used SEER 17 registries research data [(2000–2020); version 8.4.2]. Since HER-2 status was included from 2010, the inclusion criteria were as follows: (1) female; (2) year of diagnosis from 2010 to 2020; (3) histologic type ICD-O-3 of 8480/3; (4) complete clinical information; and (5) survival time greater than 1 month. Exclusion criteria was patients with multiple primary tumors. In addition, we retrospectively collected MBC patients who were treated at Jiangmen Central Hospital (JCH) (n = 98) and Cancer Hospital of Shantou University Medical College (CHSU) (n = 85) between January 2010 and October 2020 and met the inclusion criteria. This study was approved by the Ethics Committees of JCH (No. 2023146) and CHSU (No. 2023130), and was conducted in accordance with the 1964 Helsinki Declaration and its subsequent amendments, or comparable ethical standards. Our Ethics Committees granted a waiver of informed consent.

Data collection

The following patient characteristics were obtained: age, race, marital status, median household income, tumor location, histologic grade, molecular subtype, T stage, N stage, M stage, surgery, radiotherapy, and chemotherapy. The primary endpoint was OS and the secondary endpoint was BCSS .

Feature selection, model construction and evaluation

To reduce redundant variables, we used univariate and multivariate Cox regression analyses to screen for independent factors associated with prognosis. Statistically significant factors were used as features to participate in the construction of ML models. Ten common ML algorithms are used to develop models to predict OS and BCSS in 3-, 5-, and 7- year including XGBoost, logistic regression (LR), light gradient boosting machine (LightGBM), random forest (RF), adaptive boosting (AdaBoost), gaussian naive bayes (GNB), complement naive bayes (CNB), multi-layer perceptron neural networks (MLP), support vector machine (SVM), and k-nearest neighbors (KNN). To enhance the robustness of models, we tested and tuned them using ten-fold cross-validation and grid search methods to determine the optimal combination of hyperparameters. Patients from the SEER database were randomized into training and internal test groups in a 7:3 ratio. Two cohorts from Chinese hospitals were used as external test groups to further validate the extensiveness of the optimal model.

The performance of the machine learning model was evaluated by the AUC ^[20], accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F1 score. The confusion matrix visualized the classification performance of the model. Decision curve analysis (DCA) was used to determine the clinical usefulness of the model. The SHapley Additive exPlanations (SHAP) values were calculated through the "shap" package to explain the contribution or importance of each feature to the model.

We have developed an online website based on the Streamlit application to make our optimal prediction models available to clinicians.

PSM

To further compare the impact of mastectomy and BCS on the prognosis of MBC patients, we included 5760 patients from the SEER cohort. Inclusion criteria were (1) staging T1-2N0M0 and (2) undergoing mastectomy or BCS. Exclusion criteria were (1) patients who underwent mastectomy with radiotherapy and (2) patients who underwent BCS without radiotherapy. We used 1:1 PSM to balance the cohort baseline according to the features of the ML model to reduce confounding bias in retrospective studies. Univariate and multivariate Cox regression analyses were performed before and after PSM. In addition, the forest plot was used to demonstrate the survival differences between surgical approaches in different strata of MBC patients among the PSM-adjusted population.

Statistical analysis

Univariate and multivariate Cox regression analyses were used to identify the features of the modelled. Statistical analysis was performed using R software 4.2.1 (r-project.org/) or Python version 3.8 (Python Software Foundation). P < 0.05 indicates that the the difference is statistically significant.

Clinicopathologic characteristics

Finally, we collected 7553 eligible MBC patients from the SEER database. As shown in Table 1, 1257 (16.64%) patients were less than or equal to 50 years old, 2252 (29.82%) were between 51 and 65 years old, and 4044 (53.54%) were greater than or equal to 66 years old. Race was white in the majority of patients (74.22%). Half of the patients were married (49.36%) and 16.62% were single or homosexual. In terms of household income, 75.73% of patients had an income of more than $60,000 dollars. Tumors were located in the upper outer quadrant in 25.27% of patients, in the lower inner quadrant in 10.31%, in the lower outer quadrant in 9.02%, and in the central quadrant in 6.88%. Grade Ⅰ accounted for 54.84% of the patients, whereas Grade Ⅲ and Ⅳ accounted for only 8.78% of the patients. The subtype HR+/HER2- accounted for the majority of patients (94.03%).The proportions of stage T1 to T4 were 63.55%, 29.55%, 5.20% and 1.69% respectively.The proportions of patients with stage N0 to N3 were 90.67%, 7.70%,0.98% and 0.65% respectively. Only 1.22% of patients developed distant metastases. In terms of treatment, 94.55% of patients received mastectomy or BCS, 51.95% received radiotherapy and only 12.41% received chemotherapy. Furthermore, we conducted an analysis of the correlations between the variables and generated a heatmap, demonstrating the absence of multicollinearity among the variables (Fig S1).

Table 1

Baseline characteristics of patients with mucinous breast cancer in the SEER database
Variables		Cases	%
Age	≤ 50	1257	16.64
	51–65	2252	29.82
	≥ 66	4044	53.54
Race	White	5606	74.22
	Black	901	11.93
	Others	1046	13.85
Marital status	Singled/homosexual	1255	16.62
	Married	3728	49.36
	Widow/divorced/others	2570	34.03
Median household income (inflation adjusted)	<$40,000	208	2.75
	$40,00–59,999	1625	21.52
	$60,000+	5720	75.73
Tumor location	Upper outer	1909	25.27
	Lower outer	779	10.31
	Lower inner	681	9.02
	Upper inner	1026	13.58
	Central	520	6.88
	Others	2638	34.93
Grade	Well differentiated	4142	54.84
	Moderate differentiated	2748	36.38
	Poorly differentiated	193	2.56
	Unknown	470	6.22
Subtype	HR+/HER2+	370	4.9
	HR+/HER2-	7102	94.03
	HR-/HER2+	50	0.66
	HR-/HER2-	31	0.41
T stage	T1	4800	63.55
	T2	2232	29.55
	T3	393	5.20
	T4	128	1.69
N stage	N0	6848	90.67
	N1	582	7.70
	N2	74	0.98
	N3	49	0.65
M stage	M0	7461	98.78
M stage	M1	92	1.22
Surgery	No	336	4.45
	Mastectomy	2129	28.19
	Breast-conserving surgery	5088	67.36
Radiotherapy	No/unknown	3629	48.05
Radiotherapy	Yes	3924	51.95
Chemotherapy	No/unknown	6616	87.59
Chemotherapy	Yes	937	12.41

Feature selection

As shown in Table 2, univariate Cox regression analysis revealed that age, race, marital status, median household income, subtype, T stage, N stage, M stage, surgery, radiotherapy, and chemotherapy significantly affected OS. Meanwhile, significantly affecting BCSS were age, race, marital status, median household income, grade, subtype, T stage, N stage, M stage, surgery, radiotherapy, and chemotherapy.

Table 2

Univariate and multivariate Cox analyses of patients with mucinous breast cancer in the SEER database
	Univariate Cox analysis						Multivariate Cox analysis
	OS			BCSS			OS			BCSS
	HR	95%CI	P	HR	95%CI	P	HR	95%CI	P	HR	95%CI	P
Age
≤ 50	Reference			Reference			Reference			Reference
51–65	1.70	1.18–2.44	0.004	0.80	0.48–1.34	0.402	1.91	1.32–2.78	0.001	1.02	0.59–1.76	0.954
66+	7.01	5.08–9.67	< 0.001	1.64	1.06–2.52	0.025	6.36	4.50–8.99	< 0.001	2.46	1.45–4.19	0.001
Race
White	Reference			Reference			Reference			Reference
Black	1.07	0.88–1.30	0.485	1.59	1.08–2.34	0.018	0.96	0.78–1.17	0.683	1.11	0.73–1.68	0.622
Others	0.57	0.45–0.72	0.001	0.55	0.31–0.98	0.041	0.78	0.61-1.00	0.054	0.79	0.44–1.42	0.432
Marital status
Singled/homosexual	Reference			Reference			Reference			Reference
Married	0.67	0.54–0.83	< 0.001	0.45	0.29–0.69	< 0.001	0.61	0.49–0.76	< 0.001	0.60	0.38–0.95	0.029
Widow/divorced/others	1.98	1.63–2.40	< 0.001	1.26	0.85–1.85	0.247	1.05	0.86–1.29	0.632	1.04	0.67–1.61	0.857
Median household income (inflation adjusted)
<$40,000	Reference			Reference			Reference			Reference
$40,00–59,999	0.81	0.57–1.13	0.215	0.58	0.30–1.15	0.121	0.76	0.54–1.07	0.115	0.44	0.22–0.88	0.020
$60,000+	0.61	0.44–0.85	0.003	0.45	0.24–0.86	0.016	0.65	0.46–0.90	0.011	0.33	0.17–0.64	0.001
Tumor location
Upper outer	Reference			Reference			Reference			Reference
Lower outer	0.88	0.68–1.14	0.330	0.81	0.45–1.44	0.470	/	/	/	/	/	/
Lower inner	0.93	0.72–1.21	0.578	0.75	0.40–1.39	0.359	/	/	/	/	/	/
Upper inner	1.05	0.84–1.31	0.663	1.10	0.69–1.773	0.688	/	/	/	/	/	/
Central	1.15	0.88–1.51	0.317	0.72	0.35–1.46	0.360	/	/	/	/	/	/
Others	1.15	0.97–1.36	0.104	1.04	0.72–1.51	0.836	/	/	/	/	/	/
Grade
Well differentiated	Reference			Reference			Reference			Reference
Moderate differentiated	0.92	0.79–1.06	0.241	1.49	1.07–2.09	0.018	/	/	/	1.48	1.05–2.08	0.026
Poorly differentiated	0.96	0.63–1.45	0.832	3.20	1.70–6.04	< 0.001	/	/	/	2.06	1.02–4.18	0.045
Unknown	1.13	0.91–1.41	0.273	2.51	1.62–3.89	< 0.001	/	/	/	1.08	0.67 ~ 1.74	0.745
Subtype
HR+/HER2+	Reference			Reference			Reference			Reference
HR+/HER2-	1.49	1.05–2.12	0.027	0.79	0.43–1.46	0.453	0.84	0.58–1.22	0.358	1.06	0.55–2.05	0.854
HR-/HER2+	1.22	0.48–3.14	0.677	1.41	0.31–6.36	0.656	1.26	0.48–3.26	0.640	0.80	0.16–3.94	0.781
HR-/HER2-	3.37	1.55–7.31	0.002	4.78	1.52–15.01	0.007	1.77	0.81–3.86	0.154	4.58	1.38–15.17	0.013
T stage
T1	Reference			Reference			Reference			Reference
T2	1.76	1.52–2.03	< 0.001	3.03	2.10–4.35	< 0.001	1.78	1.53–2.07	< 0.001	2.14	1.45–3.16	< 0.001
T3	3.00	2.41–3.73	< 0.001	8.41	5.44–13.02	< 0.001	2.24	1.73–2.89	< 0.001	2.49	1.47–4.25	0.001
T4	5.69	4.24–7.65	< 0.001	26.08	16.36–41.56	< 0.001	3.02	2.04–4.46	< 0.001	2.68	1.35–5.34	0.005
N stage
N0	Reference			Reference			Reference			Reference
N1	1.32	1.07–1.64	0.011	3.83	2.67–5.50	< 0.001	1.17	0.92–1.50	0.208	1.63	1.04–2.54	0.031
N2	1.73	1.04–2.88	0.035	6.93	3.52–13.64	< 0.001	2.19	1.27–3.78	0.005	2.67	1.24–5.75	0.012
N3	2.96	1.74–5.02	< 0.001	16.3	8.99–29.54	< 0.001	0.65	0.34–1.25	0.194	1.23	0.55–2.75	0.612
M stage
M0	Reference			Reference			Reference			Reference
M1	8.18	6.2-10.78	< 0.001	35.09	24.54–50.16	< 0.001	2.38	1.68–3.38	< 0.001	4.27	2.61–6.98	< 0.001
Surgery
No	Reference			Reference			Reference			Reference
Mastectomy	0.15	0.13–0.19	< 0.001	0.06	0.04–0.09	< 0.001	0.28	0.23–0.36	< 0.001	0.12	0.07 ~ 0.19	< 0.001
Breast-conserving surgery	0.13	0.11–0.15	< 0.001	0.04	0.03–0.05	< 0.001	0.36	0.28–0.45	< 0.001	0.13	0.08–0.21	< 0.001
Radiotherapy
No/unknown	Reference			Reference			Reference			Reference
Yes	0.35	0.31–0.41	< 0.001	0.44	0.33–0.61	< 0.001	0.47	0.40–0.56	< 0.001	0.83	0.57–1.21	0.329
Chemotherapy
No/unknown	Reference			Reference			Reference			Reference
Yes	0.52	0.41–0.66	< 0.001	2.19	1.57–3.07	< 0.001	0.71	0.54–0.95	0.021	1.45	0.93–2.26	0.101
Abbreviations: OS, overall survival; BCSS, breast cancer-specific survival; HR, hazard ratio; CI, confidence internal.

In addition, multivariate Cox regression analysis was used to further explore independent risk factors for prognosis. Higher age, higher T stage, N3 stage, and M1 stage suggested poorer OS. Regarding social factors, being married and household income of more than $60,000 had better OS. On the other hand, undergoing surgery, radiotherapy, and chemotherapy suggested better OS. Age > = 66 years, higher grade (II and III), HR-/HER2-, higher T stage, N2-3 stage, and M1 stage suggested poorer BCSS. Whereas being married, higher household income and undergoing surgery suggested better BCSS.

Establishment and evaluation of prognostic models

We incorporated significant features into subsequent ML model construction to predict OS and BCSS of MBC patients at 3-, 5-, and 7-year. Table 3 demonstrated the performance of ten ML models in the training and internal test groups. The XGBoost models excelled in predicting OS at 3-year (training group: AUC = 0.833; internal test group: AUC = 0.839), 5-year (training group: AUC = 0.856; internal test group: AUC = 0.816), and 7-year (train set AUC = 0.843; internal test group: AUC = 0.830). Also, the XGBoost models performed excellently in BCSS at 3-year (training group: AUC = 0.944; internal test group: AUC = 0.872), 5-year (training group: AUC = 0.905; internal test group: AUC = 0.908), and 7-year (training group: AUC = 0.907; internal test group: AUC = 0.905). Other ML models such as LR (3-year OS: AUC = 0.828; 5-year OS: AUC = 0.791; 7-year OS: AUC = 0.816; 3-year BCSS: AUC = 0.847; 5-year BCSS: AUC = 0.878; 7-year BCSS: AUC = 0.913), LightGBM (3-year OS: AUC = 0.648; 5-year OS: AUC = 0.554; 7-year OS: AUC = 0.546; 3-year BCSS: AUC = 0.763; 5-year BCSS: AUC = 0.752; 7-year BCSS: AUC = 0.752), RF (3-year OS: AUC = 0.799; 5-year OS: AUC = 0.773; 7-year OS: AUC = 0.777; 3-year BCSS: AUC = 0.862; 5-year BCSS: AUC = 0.869; 7-year BCSS: AUC = 0.841), GNB (3-year OS: AUC = 0.819; 5-year OS: AUC = 0.793; 7-year OS: AUC = 0.811; 3-year BCSS: AUC = 0.838; 5-year BCSS: AUC = 0.865; 7-year BCSS: AUC = 0.812), CNB (3-year OS: AUC = 0.792; 5-year OS: AUC = 0.754; 7-year OS: AUC = 0.788; 3-year BCSS: AUC = 0.818; 5-year BCSS: AUC = 0.827; 7-year BCSS: AUC = 0.847), MLP (3-year OS: AUC = 0.583; 5-year OS: AUC = 0.515; 7-year OS: AUC = 0.805; 3-year BCSS: AUC = 0.515; 5-year BCSS: AUC = 0.598; 7-year BCSS: AUC = 0.603), SVM (3-year OS: AUC = 0.561; 5-year OS: AUC = 0.608; 7-year OS: AUC = 0.554; 3-year BCSS: AUC = 0.600; 5-year BCSS: AUC = 0.813; 7-year BCSS: AUC = 0.821) and KNN (3-year OS: AUC = 0.643; 5-year OS: AUC = 0.692; 7-year OS: AUC = 0.774; 3-year BCSS: AUC = 0.700; 5-year BCSS: AUC = 0.769; 7-year BCSS: AUC = 0.820) performed slightly worse than XGBoost and AdaBoost models (3-year OS: AUC = 0.847; 5-year OS: AUC = 0.813; 7-year OS: AUC = 0.830; 3-year BCSS: AUC = 0.865; 5-year BCSS: AUC = 0.870; 7-year BCSS: AUC = 0.903) in the internal test group.

Table 3

Performance of machine learning prognostic models in the training and internal test groups
Groups	Model performance	XGB	LR	LightGBM	RF	AdaBoost	GNB	CNB	MLP	SVM	KNN
Training group	3-year OS	0.833	0.801	0.618	0.908	0.822	0.803	0.779	0.521	0.564	0.787
	5-year OS	0.856	0.819	0.651	0.907	0.838	0.813	0.789	0.502	0.506	0.829
	7-year OS	0.843	0.795	0.666	0.896	0.818	0.791	0.762	0.646	0.568	0.825
	3-year BCSS	0.948	0.873	0.791	0.976	0.916	0.876	0.856	0.584	0.858	0.938
	5-year BCSS	0.905	0.864	0.744	0.976	0.914	0.873	0.805	0.568	0.849	0.912
	7-year BCSS	0.907	0.829	0.684	0.967	0.874	0.883	0.749	0.581	0.821	0.894
Internal test group	3-year OS	0.839	0.828	0.648	0.799	0.847	0.819	0.792	0.583	0.561	0.643
	5-year OS	0.816	0.791	0.554	0.773	0.813	0.793	0.754	0.515	0.608	0.692
	7-year OS	0.830	0.816	0.546	0.777	0.830	0.811	0.788	0.805	0.554	0.774
	3-year BCSS	0.896	0.847	0.763	0.862	0.865	0.838	0.818	0.515	0.600	0.700
	5-year BCSS	0.908	0.878	0.752	0.869	0.870	0.865	0.827	0.598	0.813	0.769
	7-year BCSS	0.905	0.913	0.752	0.841	0.903	0.812	0.847	0.603	0.821	0.820
Abbreviations: AUC, area under the curve; XGBoost, extreme gradient boosting; LR, logistic regression; LightGBM, light gradient boosting machine; RF, random forest; AdaBoost, adaptive boosting; GNB, gaussian naive bayes; CNB, complement naive bayes; MLP, multi-layer perceptron neural networks; SVM, support vector machine; KNN, k-nearest neighbors; OS, overall survival; BCSS, breast cancer-specific survival.

To further select the optimal model and validate the robustness and applicability of models, we collected a total of 183 MBC patients from JCH and CHSU cohorts (Table S1). Obviously, the XGBoost models (3-year OS: AUC = 0.889; 5-year OS: AUC = 0.889; 7-year OS: AUC = 0.884; 3-year BCSS: AUC = 0.911; 5-year BCSS: AUC = 0.856; 7-year BCSS: AUC = 0.871) performed slightly better than the AdaBoost models (3-year OS: AUC = 0.878; 5-year OS: AUC = 0.851; 7-year OS: AUC = 0.856; 3-year BCSS: AUC = 0.913; 5-year BCSS: AUC = 0.856; 7-year BCSS: AUC = 0.852) in the external test group (Fig. 2A-F). In addition, JCH and CHSU cohorts each performed equally well in both models (Fig S2). Therefore, the XGBoost models were the optimal models for predicting the prognosis of MBC patients.

Evaluation and interpretability of the XGBoost models

Table S2 listed the accuracy, sensitivity, specificity, PPV, NPV and F1 score of the ten ML models, in which the XGBoost models had an accuracy of 0.728 in predicting 3-year OS, 0.777 in 5-year OS, 0.758 in 7-year OS, 0.894 in 3-year BCSS, 0.887 in 5-year BCSS, and 0.882 in 7-year BCSS. The confusion matrix visualised the performance of the XGBoost classifiers in the internal test group (Fig S3). DCA was used to determine the clinical usefulness of the model. The results showed that the XGBoost models had a good net profit in predicting survival at different times and therefore have clinical utility (Fig. 3).

The SHAP analysis provided an explanation of how the XGBoost models predicted survival outcomes and calculated the importance of features. Figure 4A-F showed the SHAP values for each feature at different levels. As the feature values increase, the redder colour is, and vice versa, the bluer colour is. In addition, we ranked the features of the models (Fig. 4G-L). The higher feature ranking indicated that the feature is more important, which means that the feature contributes more to the model. Overall, radiotherapy, T stage, and age were the most valued by the XGBoost models for predicting 3, 5, and 7-year OS. Meanwhile, surgery, T stage, and M stage were highly valued in the XGBoost models for predicting 3-, 5-, and 7-year BCSS.

Web application development

To enhance the adoption of our prognostic models among researchers and clinicians, we designed an interactive web application using the Streamlit platform. This user-friendly application enables users to effortlessly predict the the probability of survival and survival status at various time points by inputting the clinicopathological characteristics of MBC patients (Fig. 5; https://zqc-mbc-survival.streamlit.app/). This interactive tool effectively simplifies the process of utilizing our prognostic models, making them easily accessible and valuable for users in the research and clinical communities. By developing this tool, we aim to enhance accessibility and usability, enabling users to quickly and efficiently access prognostic information for MBC patients.

Prognostic impact of surgical approaches in MBC patients

4855 MBC patients meeting criteria were used to investigate the impact of mastectomy versus BCS on patient survival outcomes. Prior to adjusting for baseline characteristics, both univariate and multivariate Cox regression analyses indicated that patients undergoing BCS had a superior OS compared to those who underwent mastectomy. However, the BCSS for patients treated with BCS did not significantly differ from those treated with mastectomy (Table S3).

PSM was used to adjust for the observed imbalance between the two groups. There were no significant differences in baseline characteristics after PSM adjustment (Table 4). After adjusting for baseline characteristics, we observed a significant 40% reduction in the overall risk of death for patients who underwent BCS (Table 5, p < 0.001, HR: 0.60, 95% confidence internal (CI): 0.47–0.77). This finding was further corroborated by multivariate Cox regression analyses. Conversely, there was no significant difference detected in the risk of BC-related death (p = 0.279, HR: 0.62, 95% CI: 0.26–1.48). To delve deeper into the disparity in OS between mastectomy and BCS across various subgroups, the forest plot revealed that BCS conferred a survival advantage in patients who were aged 66 or older, white, divorced, with a household income exceeding $40,000, grade I, HR+/HER2-, T1 and T2, and non-chemotherapy (Fig. 6).

Table 4

Comparison of patient characteristics according to surgical approaches before and after propensity score matching
Variables	Before PSM				After PSM
Variables	n	Mastectomy	BCS	P	n	Mastectomy	BCS	P
Age
≤ 50	840	379 (25.02)	461 (13.80)	< 0.001	735	379 (25.02)	356 (23.50)	0.291
51–65	1695	436 (28.78)	1259 (37.69)		910	436 (28.78)	474 (31.29)
≥ 66	2320	700 (46.20)	1620 (48.50)		1385	700 (46.20)	685 (45.21)
Race
White	3578	1050 (69.31)	2528 (75.69)	< 0.001	2105	1050 (69.31)	1055 (69.64)	0.962
Black	550	179 (11.82)	371 (11.11)		359	179 (11.82)	180 (11.88)
Others	727	286 (18.88)	441 (13.20)		566	286 (18.88)	280 (18.48)
Marital status
Singled/homosexual	794	246 (16.24)	548 (16.41)	0.264	476	246 (16.24)	230 (15.18)	0.725
Married	2586	785 (51.82)	1801 (53.92)		1578	785 (51.82)	793 (52.34)
Widow/divorced/others	1475	484 (31.95)	991 (29.67)		976	484 (31.95)	492 (32.48)
Median household income (inflation adjusted)
<$40,000	134	60 (3.96)	74 (2.22)	< 0.001	113	60 (3.96)	53 (3.50)	0.692
$40,00–59,999	1049	344 (22.71)	705 (21.11)		702	344 (22.71)	358 (23.63)
$60,000+	3672	1111 (73.33)	2561 (76.68)		2215	1111 (73.33)	1104 (72.87)
Grade
Well differentiated	2732	781 (51.55)	1951 (58.41)	< 0.001	1570	781 (51.55)	789 (52.08)	0.225
Moderate differentiated	1765	600 (39.60)	1165 (34.88)		1191	600 (39.60)	591 (39.01)
Poorly differentiated	107	46 (3.04)	61 (1.83)		77	46 (3.04)	31 (2.05)
Unknown	251	88 (5.81)	163 (4.88)		192	88 (5.81)	104 (6.86)
Subtype
HR+/HER2+	210	85 (5.61)	125 (3.74)	< 0.001	139	85 (5.61)	54 (3.56)	0.053
HR+/HER2-	4594	1403 (92.61)	3191 (95.54)		2848	1403 (92.61)	1445 (95.38)
HR-/HER2+	32	18 (1.19)	14 (0.42)		27	18 (1.19)	9 (0.59)
HR-/HER2-	19	9 (0.59)	10 (0.30)		16	9 (0.59)	7 (0.46)
T stage
T1	3413	894 (59.01)	2519 (75.42)	< 0.001	1785	894 (59.01)	891 (58.81)	0.912
T2	1442	621 (40.99)	821 (24.58)		1245	621 (40.99)	624 (41.19)
Chemotherapy
Unknown	4437	1345 (88.78)	3092 (92.57)	< 0.001	2719	1345 (88.78)	1374 (90.69)	0.083
Yes	418	170 (11.22)	248 (7.43)		311	170 (11.22)	141 (9.31)
Abbreviations: PSM, propensity score matching; BCS, breast-conserving surgery.

Table 5

Univariate and multivariate Cox analyses in patients with mucinous breast cancer after propensity score matching
	Univariate Cox analysis						Multivariate Cox analysis
	OS			BCSS			OS			BCSS
	HR	95%CI	P	HR	95%CI	P	HR	95%CI	P	HR	95%CI	P
Age
≤ 50	Reference			Reference			Reference			Reference
51–65	3.25	1.57–6.75	0.002	0.00	0-infinity	0.996	3.09	1.48–6.44	0.003	0.00	0-infinity	0.997
66+	13.58	6.97–26.45	< 0.001	3.58	1.06–12.09	0.04	11.90	5.99–23.64	< 0.001	4.61	1.27–16.69	0.02
Race
White	Reference			Reference			Reference			Reference
Black	1.00	0.68–1.46	0.997	2.26	0.82–6.29	0.117	/	/	/	/	/	/
Others	0.73	0.51–1.05	0.091	0.89	0.26–3.09	0.852	/	/	/	/	/	/
Marital status
Singled/homosexual	Reference			Reference			Reference			Reference
Married	0.77	0.52–1.15	0.201	0.28	0.09–0.86	0.027	0.66	0.44–0.99	0.043	0.22	0.07–0.69	0.01
Widow/divorced/others	1.96	1.34–2.86	< 0.001	0.78	0.28–2.14	0.624	1.06	0.72–1.56	0.774	0.36	0.12–1.05	0.062
Median household income (inflation adjusted)
<$40,000	Reference			Reference			Reference			Reference
$40,000–59,999	1.05	0.59–1.89	0.859	0.39	0.08–2.03	0.265	/	/	/	/	/	/
$60,000+	0.68	0.39–1.20	0.188	0.43	0.10–1.87	0.26	/	/	/	/	/	/
Grade
Well differentiated	Reference			Reference			Reference			Reference
Moderate differentiated	0.83	0.64–1.08	0.166	2.93	1.17–7.34	0.022	0.90	0.69–1.18	0.446	3.15	1.25–7.95	0.015
Poorly differentiated	0.66	0.27–1.61	0.364	0.00	0-infinity	0.997	1.07	0.41–2.75	0.894	0.00	0-infinity	0.999
Unknown	0.52	0.31–0.88	0.016	1.50	0.31–7.24	0.614	0.59	0.35–1.01	0.052	1.83	0.38–8.87	0.451
Subtype
HR+/HER2+	Reference			Reference			Reference			Reference
HR+/HER2-	2.31	0.95–5.59	0.064	3.05E + 07	0-infinity	0.998	1.46	0.58–3.71	0.422	/	/	/
HR-/HER2+	0.88	0.10–7.54	0.908	1.00	0-infinity	1	0.67	0.08–6.05	0.724	/	/	/
HR-/HER2-	6.74	1.61–28.22	0.009	3.77E + 08	0-infinity	0.997	2.60	0.61–11.09	0.197	/	/	/
T stage
T1	Reference			Reference			Reference			Reference
T2	1.65	1.30–2.11	< 0.001	2.41	1.03–5.64	0.043	1.58	1.24–2.02	< 0.001	2.24	0.95–5.25	0.064
Surgery
Mastectomy		Reference			Reference			Reference			Reference
Breast-conserving surgery	0.60	0.47–0.77	< 0.001	0.60	0.25–1.43	0.249	0.60	0.47–0.78	< 0.001	0.62	0.26–1.48	0.279
Chemotherapy
No/unknown	Reference			Reference			Reference			Reference
Yes	0.45	0.26–0.79	0.005	1.94	0.66–5.73	0.231	1.09	0.58–2.08	0.786	/	/	/
Abbreviations: OS, overall survival; BCSS, breast cancer-specific survival; HR, hazard ratio; CI, confidence internal.

MBC, as a rare histological type, has received less attention due to its relatively favorable prognosis^{[21, 22]}. Most molecular subtypes of MBC present as ER+/HER2-, and the treatment for MBC patients usually follows the guidelines for IDC, with a focus on surgery, chemotherapy, and endocrine therapy^[23]. However, Pareja et al. have revealed through genomic landscape analysis that MBC is a genetically heterogeneous form of BC that is distinct from other common ER+/HER2-^[7]. Therefore, it is highly necessary to personalize the treatment and prognosis prediction for MBC. To our knowledge, our study is the largest one to date in analyzing the prognosis and surgical approach of MBC patients. This study is the first to develop models for predicting OS and BCSS based on ten ML algorithms. Our XGBoost models demonstrated superior sensitivity, specificity, and accuracy in predicting 3-, 5-, and 7-year OS and BCSS in MBC patients. In addition, we were the first to explore the survival benefit of mastectomy and BCS for MBC patients by PSM.

We identified several independent risk factors significantly associated with both OS and BCSS, including age > = 66 years, higher T stage, N2 stage and M1 stage. Whereas, independent protective factors included being married, household income of more than $60,000 and undergoing surgery. Several recent studies have shown that older age tends to imply poorer OS and BCSS for patients, with cut-off values for age including 52, 65 and 80 years^{[13, 15, 24]}. In general, higher malignancy stage represents worse prognosis for patients, and our results similarly indicated that higher TNM stage serves as a risk indicator for MBC prognosis. Numerous studies have highlighted marital status as a significant predictor of survival in BC patients^[25–29], with married patients exhibiting a better quality of life and improved survival compared to unmarried or divorced patients^[30]. Certainly, high income families are more likely to adhere to clinicians' advice and recommendations, thus benefiting from enhanced decision-making in therapeutic regimens due to the absence of financial strain^{[31, 32]}. Consequently, our research revealed that patients with family income of more than $60,000 demonstrated improved prognoses. Extensive studies have shown that patients undergoing mastectomy or BCS generally experience better outcomes than those who forgo surgical intervention, owing to the reduction of primary tumor load^[33–36], aligning with our results. In addition, radiotherapy and chemotherapy are independent factors influencing OS but not BCSS in MBC patients. Mo et al. identified radiotherapy as an independent factor for BCSS in MBC patients with T1-2N0M0 (T ≤ 3 cm)^[37]. It is possible that radiotherapy may confer survival advantages in a specific subgroup of MBC patients, but since our study focused on the overall MBC population, no statistically significant observations were made regarding this issue. Previous research has reported that patients receiving chemotherapy exhibit enhanced OS in comparison to those who do not after PSM, yet this did not translate into a marked difference in BCSS^[38], consistent with our findings.

In our comparative analysis of ten ML models, the XGBoost algorithm emerged as the top-performing model, surpassing all other models in a large American cohort and consistently validated in cross-national cohorts. Fu et al. previously developed a nomogram for 5- and 7-year BCSS in early MBC patients, achieving a concordance index (C-index) of 0.789^[14]. In contrast, our XGBoost models demonstrated remarkable predictive power for 5- and 7-year BCSS, with AUC values of 0.905 and 0.907, respectively, in the training group. Furthermore, when applied to the external test group, the model sustained its robust performance, yielding AUCs of 0.856 and 0.871, respectively. Zhu et al. established a prognostic nomogram to predict 3- and 5-year OS in MBC patients with a concordance index (C-index) of 0.803^[15]. Similarly, Gao et al. devised a nomogram aimed at predicting 5- and 10-year OS in MBC patients, reporting AUC values of 0.714, 0.813, and 0.805 for 5-year OS across the training, internal validation, and external validation cohorts^[13]. In comparison, our XGBoost models demonstrated superior predictive performance, with AUC values of 0.833, 0.839, and 0.889 for 3-year OS in these groups, along with AUC values of 0.856, 0.816, and 0.889 for 5-year OS, respectively. Our findings revealed that the predictive performance of our XGBoost models greatly surpassed that of earlier nomograms, indicating the enhanced efficacy of our XGBoost model in prognosticating OS and BCSS for MBC patients. This enhancement in predictive performance may offer a more reliable evidence for clinicians in managing and stratifying MBC patients. Meanwhile, the DCA curve confirmed the exceptional clinical utility of our XGBoost model. To facilitate clinical application, we have developed an accessible online web tool that allows clinicians to rapidly estimate the survival probabilities for individual MBC patients.

Since the landmark NSABP B-06 trial, it has been established that the survival of early BC patients undergoing BCS are comparable to those of patients undergoing mastectomy^[39]. Two subsequent large-scale studies have demonstrated superior survival outcomes for patients with early BC who underwent BCS with radiotherapy, as compared to those who underwent mastectomy without radiotherapy^{[40, 41]}. Therefore, clinicians are increasingly inclined to recommend BCS with radiotherapy over mastectomy for patients who meet the indication for BCS. However, the survival benefit of BCS with radiotherapy versus mastectomy in MBC patients remains unproven. In light of this, our study included MBC patients at stage T1-2N0M0 and employed a PSM method to mitigate confounding effects, thereby simulating the randomization of survival benefits between the BCS and mastectomy groups. Our findings revealed that after PSM, the OS of the BCS group was significantly superior to that of the mastectomy group (p < 0.001, HR = 0.60, 95% CI: 0.47–0.78). In contrast, the BCSS in the BCS group did not significantly differ from that in the mastectomy group (p = 0.279, HR = 0.62, 95% CI: 0.26–1.48). These results align with the findings of Yu et al^[42], despite their study not employing the PSM method to adjust for potential confounding bias. Thus, our study provides robust evidence that MBC patients with stage T1-2N0M0 are more suitable to receive BCS with radiotherapy for achieving better prognosis.

Nevertheless, it is important to acknowledge the limitations of our study. Firstly, our study was retrospective, potentially susceptible to selection bias and other confounding factors that warrant validation in a large prospective cohort. Secondly, the SEER database does not incorporate information on endocrine therapy and targeted therapy, both of which have a significant impact on patient prognosis. This missing data would degrade model performance. Finally, the omission of data on endocrine therapy in the SEER database led to the exclusion of older patients with stage T1 who underwent BCS and received endocrine therapy without radiotherapy. This exclusion potentially introduces a selection bias into the prognostic comparison between mastectomy and BCS.

In summary, we developed six optimal prognostic models utilizing the XGBoost algorithm to predict the survival of MBC patients. External validation confirmed the high generalizability of our models. Notably, we observed a significant improvement in OS for patients undergoing BCS.

AdaBoost, adaptive boosting; AUC, area under the curve; BC, breast cancer; BCS, breast-conserving surgery; BCSS, breast cancer-specific survival; CHSU, Cancer Hospital of Shantou University Medical College; CI, confidence internal; C-index, concordance index; CNB, complement naive bayes; DCA, decision curve analysis; GNB, gaussian naive bayes; IDC, infiltrating ductal carcinoma; JCH, Jiangmen Central Hospital; KNN, k-nearest neighbors; LightGBM, light gradient boosting machine; LR, logistic regression; MBC, mucinous breast cancer; ML, machine learning; MLP, multi-layer perceptron neural networks; NPV, negative predictive value; OS, overall survival; PPV, positive predictive value; PSM, propensity score matching; RF, random forest; SEER, Surveillance Epidemiology and End Results; SHAP, SHapley Additive exPlanations; SVM, support vector machine; XGBoost, extreme gradient boosting.

Conflict of interests

The authors declare that there is no conflict of interest regarding the publication of this paper.

Ethical approval

This study was approved by the Ethics Committees of Jiangmen Central Hospital (No. 2023146) and Cancer Hospital of Shantou University Medical College (No. 2023130), and a waiver of informed consent was granted.

Funding

This work was supported by the Youth Science Foundation of Jiangmen Central Hospital [Grant No. J202404].

Author Contribution

Conceptualization, QZ, CC and YL; methodology, QZ, CC and YL; formal analysis, CZ, ZY and QZ; data curation, MS, YD, RL and YF; writing—original draft preparation, QZ, CC and YL; writing—review and editing, QZ, CC and YL; supervision, JW, BX and WL. All authors have read and agreed to the published version of the manuscript.

Data Availability

The datasets for this study were obtained from the SEER database [https://seer.cancer.gov/ (accessed 15 October 2023)] and two hospitals in China and are available on request to the corresponding authors.

Kaoku, S. et al. Sonographic and pathologic image analysis of pure mucinous carcinoma of the breast. Ultrasound Med. Biol. 39, 1158–1167. 10.1016/j.ultrasmedbio.2013.02.014 (2013).
Azamjah, N., Soltan-Zadeh, Y. & Zayeri, F. Global Trend of Breast Cancer Mortality Rate: A 25-Year Study. Asian Pac. J. Cancer Prev. 20, 2015–2020. 10.31557/APJCP.2019.20.7.2015 (2019).
Giaquinto, A. N. et al. Breast Cancer Stat. 2022 CA Cancer J. Clin. ;72:524–541. doi: 10.3322/caac.21754. (2022).
Lei, L., Yu, X., Chen, B., Chen, Z. & Wang, X. Clinicopathological Characteristics of Mucinous Breast Cancer: A Retrospective Analysis of a 10-Year Study. PLoS One. 11, e0155132. 10.1371/journal.pone.0155132 (2016).
Cao, A. Y. et al. Outcome of pure mucinous breast carcinoma compared to infiltrating ductal carcinoma: a population-based study from China. Ann. Surg. Oncol. 19, 3019–3027. 10.1245/s10434-012-2322-6 (2012).
Hashmi, A. A. et al. Mucinous Breast Carcinoma: Clinicopathological Comparison With Invasive Ductal Carcinoma. Cureus. 13, e13650. 10.7759/cureus.13650 (2021).
Pareja, F. et al. The Genomic Landscape of Mucinous Breast Cancer. J. Natl. Cancer Inst. 111, 737–741. 10.1093/jnci/djy216 (2019).
Roux, P. et al. Tubular and mucinous breast cancer: results of a cohort of 917 patients. Tumori. 105, 55–62. 10.1177/0300891618811282 (2019).
Diab, S. G. et al. Tumor characteristics and clinical outcome of tubular and mucinous breast carcinomas. J. Clin. Oncol. 17, 1442–1448. 10.1200/JCO.1999.17.5.1442 (1999).
Wasif, N., McCullough, A. E., Gray, R. J. & Pockaj, B. A. Influence of uncommon histology on breast conservation therapy for breast cancer-biology dictates technique. J. Surg. Oncol. 105, 586–590. 10.1002/jso.22132 (2012).
Gradishar, W. J. et al. Breast Cancer, Version 3.2022, NCCN Clinical Practice Guidelines in Oncology. J. Natl. Compr. Canc Netw. 20, 691–722. 10.6004/jnccn.2022.0030 (2022).
Di Saverio, S., Gutierrez, J. & Avisar, E. A retrospective review with long term follow up of 11,400 cases of pure mucinous breast carcinoma. Breast Cancer Res. Treat. 111, 541–547. 10.1007/s10549-007-9809-z (2008).
Gao, T. et al. Nomogram for predicting survival in patients with mucinous breast cancer undergoing chemotherapy and surgery: a population-based study. Eur. J. Med. Res. 28, 415. 10.1186/s40001-023-01395-x (2023).
Fu, J. et al. Clinical Nomogram for Predicting Survival Outcomes in Early Mucinous Breast Cancer. PLoS One. 11, e0164921. 10.1371/journal.pone.0164921 (2016).
Zhu, X. et al. Construction of a Prognostic Nomogram Model for Patients with Mucinous Breast Cancer. J. Healthc. Eng. 2022, 1230812. 10.1155/2022/1230812 (2022).
Doscher, M., Beutel, B. G., Lovy, A., Alba, B. & Melone, C. En Masse Excision and Curettage for Periarticular Gouty Tophi of the Hands. Eplasty. 22, e25 (2022).
Tran, K. A. et al. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med. 13, 152. 10.1186/s13073-021-00968-x (2021).
Yu, Y. & Tran, H. An XGBoost-Based Fitted Q Iteration for Finding the Optimal STI Strategies for HIV Patients. IEEE Trans. Neural Netw. Learn. Syst. PP 10.1109/TNNLS.2022.3176204 (2022).
Yu, P. et al. Breast-conserving therapy shows better prognosis in mucinous breast carcinoma compared with mastectomy: A SEER population-based study. Cancer Med. 9, 5381–5391. 10.1002/cam4.3202 (2020).
Obuchowski, N. A. & Bullen, J. A. Receiver operating characteristic (ROC) curves: review of methods with applications in diagnostic medicine. Phys. Med. Biol. 63, 07TR01. 10.1088/1361-6560/aab4b1 (2018).
Marrazzo, E. et al. Mucinous breast cancer: A narrative review of the literature and a retrospective tertiary single-centre analysis. Breast. 49, 87–92. 10.1016/j.breast.2019.11.002 (2020).
Sas-Korczyńska, B., Mituś, J., Stelmach, A., Ryś, J. & Majczyk, A. Mucinous breast cancer - clinical characteristics and treatment results in patients treated at the Oncology Centre in Kraków between 1952 and 2002. Contemp. Oncol. (Pozn). 18, 120–123. 10.5114/wo.2014.42727 (2014).
Lian, W., Zheng, J. & Chen, D. Different prognosis by subtype in the early mucinous breast cancer: a SEER population-based analysis. Transl Cancer Res. 9, 5969–5978. 10.21037/tcr-20-1237 (2020).
Ding, S. et al. Predictors for Survival and Distribution of 21-Gene Recurrence Score in Patients With Pure Mucinous Breast Cancer: A SEER Population-Based Retrospective Analysis. Clin. Breast Cancer. 19, e66–e73. 10.1016/j.clbc.2018.10.001 (2019).
Jiao, D. et al. Impact of Marital Status on Prognosis of Patients With Invasive Breast Cancer: A Population-Based Study Using SEER Database. Front. Oncol. 12, 913929. 10.3389/fonc.2022.913929 (2022).
Martínez, M. E. et al. Prognostic significance of marital status in breast cancer survival: A population-based study. PLoS One. 12, e0175515. 10.1371/journal.pone.0175515 (2017).
Ding, W. et al. Dynamic changes in marital status and survival in women with breast cancer: a population-based study. Sci. Rep. 11, 5421. 10.1038/s41598-021-84996-y (2021).
Buja, A. et al. Marital status and stage of cancer at diagnosis: A systematic review. Eur. J. Cancer Care (Engl). 27 10.1111/ecc.12755 (2018).
Guan, T. et al. Association of marital status with cardiovascular outcome in patients with breast cancer. J. Thorac. Dis. 14, 841–850. 10.21037/jtd-21-1261 (2022).
Kang, D. et al. Divorce after breast cancer diagnosis and its impact on quality of life. Palliat. Support Care. 20, 807–812. 10.1017/S1478951521001711 (2022).
Lehrer, S., Green, S. & Rosenzweig, K. E. Affluence and Breast Cancer. Breast J. 22, 564–567. 10.1111/tbj.12630 (2016).
Riba, L. A., Gruner, R. A., Alapati, A. & James, T. A. Association between socioeconomic factors and outcomes in breast cancer. Breast J. 25, 488–492. 10.1111/tbj.13250 (2019).
Early and locally advanced breast cancer: diagnosis and management. London. (2023).
Soran, A. et al. Randomized Trial Comparing Resection of Primary Tumor with No Surgery in Stage IV Breast Cancer at Presentation: Protocol MF07-01. Ann. Surg. Oncol. 25, 3141–3149. 10.1245/s10434-018-6494-6 (2018).
Gaitanidis, A. et al. Refusal of Cancer-Directed Surgery by Breast Cancer Patients: Risk Factors and Survival Outcomes. Clin. Breast Cancer. 18, e469–e476. 10.1016/j.clbc.2017.07.010 (2018).
Marks, C. E. et al. Metastatic breast cancer: Who benefits from surgery. Am. J. Surg. 223, 81–93. 10.1016/j.amjsurg.2021.07.018 (2022).
Mo, Q., Wang, Y., Shan, J. & Wang, X. Effect of postoperative radiotherapy in women with localized pure mucinous breast cancer after lumpectomy: a population-based study. Radiat. Oncol. 17, 119. 10.1186/s13014-022-02082-7 (2022).
Gao, H. F. et al. Adjuvant chemotherapy could benefit early-stage ER/PR positive mucinous breast cancer: A SEER-based analysis. Breast. 54, 79–87. 10.1016/j.breast.2020.09.003 (2020).
Fisher, B. et al. Eight-year results of a randomized clinical trial comparing total mastectomy and lumpectomy with or without irradiation in the treatment of breast cancer. N Engl. J. Med. 320, 822–828. 10.1056/NEJM198903303201302 (1989).
de Boniface, J., Szulkin, R. & Johansson, A. Survival After Breast Conservation vs Mastectomy Adjusted for Comorbidity and Socioeconomic Status: A Swedish National 6-Year Follow-up of 48 986 Women. JAMA Surg. 156, 628–637. 10.1001/jamasurg.2021.1438 (2021).
van Maaren, M. C. et al. 10 year survival after breast-conserving surgery plus radiotherapy compared with mastectomy in early breast cancer in the Netherlands: a population-based study. Lancet Oncol. 17, 1158–1170. 10.1016/S1470-2045(16)30067-5 (2016).
Yu, P. et al. Breast-conserving therapy shows better prognosis in mucinous breast carcinoma compared with mastectomy: A SEER population-based study. Cancer Med. 9, 5381–5391. 10.1002/cam4.3202 (2020).

No competing interests reported.

Supplementarymaterials.docx

Development and validation of novel machine learning-based prognostic models and propensity score matching for comparison of surgical approaches in mucinous breast cancer: a multicenter study

Status:

Version 1

Abstract

Figures

Introduction

Materials and methods

Patients and study design

Data collection

Feature selection, model construction and evaluation

PSM

Statistical analysis

Results

Clinicopathologic characteristics

Feature selection

Establishment and evaluation of prognostic models

Evaluation and interpretability of the XGBoost models

Web application development

Prognostic impact of surgical approaches in MBC patients

Discussion

Conclusion

Abbreviations

Declarations

Funding

Author Contribution

Data Availability

References

Additional Declarations

Supplementary Files

Status:

Version 1