Predicting diagnosis and survival of bone metastasis in breast cancer using machine learning

doi:10.21203/rs.3.rs-2978635/v1

Background

This study aimed at establishing more accurate predictive models based on novel machine learning algorithms, with the overarching goal of providing clinicians with effective decision-making assistance.

Methods

We retrospectively analyzed 283,373 breast cancer patients recorded in the Surveillance, Epidemiology, and End Results (SEER) database from 2010 to 2016. Multivariable logistic regression analyses were used to identify risk factors for bone metastases in breast cancer, whereas Cox proportional hazards regression analyses were used to identify prognostic factors for breast cancer with bone metastasis (BCBM). Based on the identified risk and prognostic factors, we developed diagnostic and prognostic models that incorporate six machine learning classifiers. We then used the area under the receiver operating characteristic (ROC) curve (AUC), learning curve, precision curve, calibration plot, and decision curve analysis to evaluate performance of the machine learning models.

Results

Univariable and multivariable logistic regression analyses showed that bone metastases were significantly associated with age, race, sex, grade, T stage, N stage, surgery, radiotherapy, chemotherapy, tumor size, brain metastasis, liver metastasis, lung metastasis, breast subtype, and PR. Univariate and multivariate Cox regression analyses revealed that age, race, marital status, grade, surgery, radiotherapy, chemotherapy, brain metastasis, liver metastasis, lung metastasis, breast subtype, ER, and PR were closely associated with the prognosis of BCBM. Among the six machine learning models, the XGBoost algorithm predicted the most accurate results (AUC = 0.98; AUC = 0.88). According to the Shapley additive explanations (SHAP), the most critical feature of the diagnostic model was surgery, followed by N stage. Interestingly, surgery was also the most critical feature of prognostic model, followed by liver metastasis.

Conclusions

Based on the XGBoost algorithm, we could effectively predict the occurrence of bone metastases in breast cancer and provide targeted references for the treatment of BCBM patients.

Health sciences/Oncology

Health sciences/Risk factors

breast cancer

bone metastasis

SEER database

predictive models

machine learning

SHAP

Currently, breast cancer (BC) is the most common malignant tumor that endangers women’s health. According to 2022 cancer statistics, BC has the highest proportion of malignancies diagnosed in American women, accounting for 31% of new cases, and is also the second leading cause of cancer-related death ¹. With the continuous improvement of BC survival rate, the number of patients with breast cancer metastasis is also increasing ^2,3. Numerous studies have shown that BC exhibits metastatic heterogeneity with distinctive metastatic precedence to diverse organs, thereby resulting in significant differences in prognoses and therapy response of BC patients ^4–6.It is well known that bone is the most common site for distant metastases of BC, with nearly 75% of distant metastasis being bone metastasis (BM) ⁷. Given the complexity of metastatic BC therapies, the treatment of BC with bone metastasis (BCBM) is limited to cytotoxic chemotherapies, endocrine therapies, and targeted therapies ⁸. Furthermore, although the 5-year overall survival rate of BC patients without metastasis is greater than 80% ⁹, distant metastases significantly reduces this rate to only about 25% ¹⁰. Strikingly, the 5-year overall survival rate for BM is even lower, at a measly 22.8% ¹¹. Studies have also revealed that bone-related events caused by BM, such as bone fracture, hypercalcemia, or spinal cord compression, have a significantly negative impact on the prognosis of BCBM patients ^12–14. Therefore, it is crucial to identify patients who may have bone metastasis and predict their survival rate. This information can guide the subsequent examination, treatment, and management of the patients' clinical outcomes.

Over the years, tumor-node-metastasis (TNM) staging system, proposed by the American Joint Committee on Cancer (AJCC), and pathological classification, proposed by the World Health Organization (WHO), have been considered as prognostic evaluation systems for BCBM ^15,16. It is worth noting that these systems only incorporate predictors such as tumor infiltrating depth, invasive site, proliferative marker, gene expression assays, and response to neoadjuvant therapy. In recent years, many prediction models for BCBM have been developed, with factors such as age, sex, race, treatment, and grade being the predictors ^17,18. However, these models have specific room for improvement in practicality and accuracy. This study aims at establishing a more accurate clinical model, with as many valid variables as possible.

With regard to model development, although nomogram is currently the most commonly used prediction model, Machine learning is favored by more and more medical workers because of its practicality, innovation and accuracy. This study, a reliable BCBM prediction model is developed through the horizontal comparison of multiple indicators based on demographic characteristics, pathological information, and survival data retrieved from the Surveillance, Epidemiology, and End Results (SEER) database. Furthermore, we developed two web pages as an extension of our research. These web pages enable clinicians to obtain precise and quantitative assessments of the likelihood of bone metastasis and the five-year survival rate for breast cancer by inputting simple data. Finally, our aim is to stratify the risk of possible bone metastasis and poor prognosis in breast cancer patients. Help clinicians make decisions, reduce the unnecessary medical burden of patients, and greatly improve the quality of patients’ life.

Study population

We collected data from on patients diagnosed with BC between 2010 and 2016 from the SEER database using the SEER*Stat software version 8.3.8.1. Notably, the SEER database, supported by the National Cancer Institute (NCI), covers about 30% of the United States population based on data collected by nearly 18 large cancer registries across the United States ¹⁹. This study did not require approval by the ethics committee, as well as patient consent and agreement because the data was publicly available and there was no specific personal information.

Variables, including age, race, sex, laterality, marital status, grade, AJCC TNM stage, surgery, radiotherapy, chemotherapy, tumor size, bone metastases, liver metastases, lung metastases, brain metastases, breast subtype, ER, PR, and HER2 were extracted from the SEER database. Patients were included according to the following criteria: (1) breast cancer confirmed by biopsy or pathology; (2) age at diagnosis ≥ 20 years; and (3) diagnosed between 2010 and 2016. The following patients were excluded: (1) patients only diagnosed via autopsy or death certificate; (2) breast cancer was not the first primary malignant tumor; and (3) cases with unknown variables. Given that machine learning needs to classify the data and set labels, we set the label of diagnostic model to be bone metastasis, and chose the 5-year survival status as the target of the prognostic model. Considering that some patients were followed for a short period of time and their exact 5-year survival status could not be obtained, we excluded patients who had been followed up for less than 5 years and survived. This was done to maintain rigor of the study and reduce possible selection bias.

Feature selection and validation strategy

To minimize the negative impact of overfitting, we performed feature selection to remove irrelevant or redundant invalid features. In the diagnostic model, univariable and multivariable logistic regression were performed to screen for risk factors. In the prognostic model, univariate and multivariate Cox proportional hazards regression were applied to screen for prognostic factors. In addition, to further optimize the model using 5-fold cross-validation, we carried out five repeated experiments on the training data and determined the best parameters for each model in the training cohort by grid search method. Finally, the importance of each feature was ranked by Shapley additive explanations (SHAP). One of the major flaws of machine learning is that it builds models similar to a ‘black box’, and thus the results are difficult to interpret. To solve this problem, SHAP was developed to analyze the impact of each feature on the predicted results.

The overall dataset collected from the SEER database was randomly divided into two cohorts in a ratio of 7:3, namely training cohort and validation cohort. Metrics such as area under the curve (AUC), accuracy, precision, recall, and F1-score were then used to evaluate the reliability of six machine learning models. Next, calibration curves were constructed and used to compare discrimination between the distinct models. Decision curve analysis (DCA) is a novel algorithm that is commonly used to estimate the net benefit value of a model under different thresholds. Compared to the evaluation indicators mentioned above, DCA could better reflect clinical efficacy of predictive models. After conducting a comprehensive comparison of diverse machine learning models, we chose the model with best prediction ability as the final predictive model. To further confirm the applicability of the selected model, it was evaluated in the validation cohort.

Machine learning algorithms

Python software was used to build machine learning predictive models. It should be noted that the scikit-learn 0.24.1 package is a very important machine learning library in Python filed, which supports four major machine learning algorithms: classification, regression, reduced dimension and clustering. It also includes three modules: feature extraction, data processing and model evaluation. This retrospective study, mainly included six common machine learning algorithms of this package.

Logistic regression is a generalized linear regression analysis model. Although the dependent variables of logistic regression can be dichotomized or multi-classified, dichotomous ones are more common and easier to explain. Logistic regression is mainly used in epidemiology to explore the risk factors of a disease and predict the probability of occurrence of a disease according to the risk factors.

The decision tree classification algorithm is an instance-based inductive learning method, which can extract a tree-like classification model from the given unordered training samples. The complexity of the predictive classification algorithm is only associated with the number of layers of the decision tree, which is linear, and the data processing efficiency is very high, which is suitable for the occasion of real-time classification. In machine learning, a decision tree is a predictive model which represents a mapping relationship between features and tags. Each node in the tree represents an object, whereas each fork path represents a possible attribute value. Finally, each leaf node corresponds to the value of the object represented by the path from the root node to the leaf node.

Random forest, as the name suggests, establishes a forest in a random way. There are many decision trees in the forest, and there is no correlation between each decision tree in the random forest. It adopts the re-sampling technique of bootstrap to repeatedly and randomly select B samples from the original training sample set with N as the training set, and the other samples as the test set.

Extra tree, an algorithm similar to random forest, uses a series of decision trees to make the final prediction of the class or category to which the data point belongs. However, the difference between extra tree and random forest is that it uses the entire original sample instead of subsampling and replacing the data like a random forest. Another difference is the way nodes are segmented. Although the random forest always chooses the best possible segmentation, the extra tree chooses random segmentation. However, both extra tree and random forest are programmed to optimize the final results.

Bayesian classifier is a general term for a class of classification algorithms, all of which are based on Bayes' theorem. The classification principle of Bayesian classifier is to use a priori probability and Bayesian formula to calculate a posteriori probability, and then select the classification result corresponding to the maximum posterior probability. This study used the Gaussian Naive Bayesian (Gaussian NB) model in the Bayesian classifier.

Extreme gradient boosting (XGBoost) model was developed by the Guestrin group in 2016. Given its fast and accurate properties, the model quickly became famous in machine learning related competitions, and is now widely used in the industrial field. It is an improvement on gradient boosting decision tree (GBDT), which has the remarkable feature of efficiently and flexibly processing missing data and assembling weak predictive models to build accurate predictive models. Notably, XGBoost is more original and better compared with traditional machine learning algorithms.

Statistical analysis

SPSS 25.0 and R 4.0.5 software were used for data description and statistical analysis. Categorical variables are expressed as percentages, whereas continuous variables are expressed as means or medians. Continuous variables conforming to normal distribution were analyzed using Student’s t-test and are presented as mean ± standard deviation. On the other hand, continuous variables not conforming to normal distribution were analyzed using the Mann-Whitney U test and are presented as median ± interquartile range. Categorical data were tested using the Chi-square test or Fisher’s exact test. Univariable and multivariable logistic regression analyses were performed to identify risk factors for BM, and variables with P < 0.05 in multivariable logistic regression analyses were finally included in the diagnostic model. Similarly, univariate and multivariate Cox proportional hazard regression analysis were performed to identify predictors of prognosis for BM, and variables with P < 0.05 in multivariate Cox proportional hazard regression analyses were finally incorporated into the prognostic model. P < 0.05 was considered statistically significant.

Population features

A total of 283373 BC patients were extracted from the SEER database (198364 patients in the training cohort and 85009 patients in the validation cohort) based on the strict inclusion and exclusion criteria, among which 3492 BCBM patients (2448 patients in the training cohort and 1044 patients in the validation cohort) were screened out (Figure 1). Table 1 shows the demographic and pathological characteristics of BC patients. It was evident that most baseline features were not significantly different between the training and validation cohorts. The results in Table 1 show that the rate of BCBM was about 2.3%, the predominated age was 40-79 years old, and the incidence rate of white people was also much higher than that of other races. Table 2 shows the baseline characteristics of BCBM patients. It was found that the median tumor size of BCBM was obviously larger (44 mm vs 18 mm), the risk of distant metastases was significantly increased, and the proportion of Luminal A BC reached 65.7% (Table 2). All features, both in the diagnostic model and the prognostic model, were analyzed by the Pearson correlation test, and the correlation heat map proved that the variables were independent of each other (Supplementary Figure 1).

Risk and prognostic factors for BCBM

Univariable logistic regression analysis demonstrated that BC patients who underwent surgery had a significantly lower risk of developing BM, suggesting that surgery was the most prominent protective factor (OR=0.023, 95%CI: 0.021-0.024) (Table 3). There were six main risk factors for BM, including grade, T stage, N stage, brain metastasis, lung metastasis, and liver metastasis. The most salient risk factor was brain metastasis (OR=86.763, 95%CI: 70.737-106.884). Furthermore, the multivariable logistic regression analysis revealed that receiving surgery was still a strong protective factor (OR=0.048, 95%CI: 0.044-0.053), and the most statistically significant risk factors were T stage (OR=6.137, 95%CI: 5.294-7.119), N stage (OR=6.648, 95%CI: 5.810-7.603) and liver metastasis (OR=9.341, 95%CI: 8.141-10.723).

Table 4 shows the statistical results of univariate and multivariate Cox regression analyses. Univariate Cox regression showed that surgery (OR=0.500, 95%CI:0.457-0.548) was a significant protective factor for the prognosis of BCBM. On the other hand, age, marital status, grade, T stage, brain metastases, lung metastases, liver metastases, and breast subtype were risk factors for the prognosis of BCBM. Moreover, the multivariate Cox regression analysis found that age, race, marital status, grade, surgery, radiotherapy, chemotherapy, brain metastases, lung metastases, liver metastases, breast subtype, ER, and PR were independent predictors of BCBM prognosis. Among them, the most prominent protective factors were surgery (OR=0.569, 95%CI: 0.517-0.628) and ER (OR=0.482, 95%CI: 0.339-0.686). In addition, advanced age, increased tumor grade, and concomitant distant metastases (brain, lung, or liver) worsened the survival outcomes of BCBM patients.

Predictive performance of the machine learning models for diagnosis and prognosis

Diagnostic model

Herein, six machine learning models were developed and evaluated through learning curves, AUC, PR curves, and calibration curves. With the continuous increase of learning samples, the learning ability of the model tended to be stable, and finally XGBoost stood out from all models (Figure 2). Figure 3 shows the evaluation curves of the six models. Considering the superiority of machine learning, the AUC values of all models exceeded 0.80, and the AUC values of XGBoost reached an astonishing 0.987 and 0.940 in the training cohort and the validation cohort, respectively. However, given that the distribution of positive and negative events in the dataset was uneven, AUC alone was not sufficient to explain the performance of the model. Therefore, the PR curve was generated to make up for the inadequacy of the receiver operating characteristic (ROC) curve, thereby further evaluating the strengths and weaknesses of the model. From Figure 3C and 3D, it is evident seen that the average precision of the accuracy of the XGBoost model was higher than that of other models. Finally, the calibration curve was drawn to compare the discrimination of each model, with results showing that the XGBoost model still maintained the best state. Table 5 summarizes the evaluation index values of all models. Collectively, these results showed that XGBoost had the most outstanding comprehensive performance, with the highest AUC (0.987), accuracy (0.947), precision (0.948), recall (0.947), and F1-score (0.947).

Prognostic model

Five-fold cross-validation was applied to evaluate the performance of the machine learning prediction model, and the results obtained after 5 repetitions and the average ROC curve of the different generated ROC curves were used as the evaluation metric. Based on the model we built, the XGBoost model performed the best in five-fold cross-validation with an average AUC of 0.79 (Figure 4). It was evident that the XGBoost model displayed excellent accuracy both in the training set and validation set. Furthermore, DCA suggested that the XGBoost model exhibited a better clinical application value in both training set and validation set, and it performed significantly better than the traditional AJCC staging system (Figure 5). In the training cohort, the XGBoost model scored the highest with an AUC of 0.880, an accuracy of 0.890, a precision of 0.870, a recall of 0.890 and a F1-score of 0.860. The XGBoost model also scored the highest in the validation cohort, with an AUC of 0.800, an accuracy of 0.880, a precision of 0.840, a recall of 0.880, and a F1-score of 0.840 (Figure 6). Finally, a heatmap was generated to indicate the prediction effect of the distinct models (Supplementary Figure 2).

Characteristic importance in the machine learning models

The SHAP diagram was utilized to more intuitively express the importance of each feature of the model. According to the multivariable logistic regression and multivariate Cox proportional hazards regression results, we included 15 and 13 features in the diagnostic and prognostic models, respectively. The SHAP plot was then used to rank these important features, indicating the degree of influence of different features on diagnosis and prognosis. From Figure 7, it is not difficult to find that the higher the SHAP value of a feature, the greater the probability of BM in BC patients. Blue indicated that the eigenvalues were small, purple indicated that the eigenvalues were close to mean value, and red indicated that the eigenvalues were large. Taking the most striking feature in the figure as an example, we found that the incidence of BM was significantly reduced in patients who underwent surgery. Figure 8 shows that surgery still remained the most important feature, with results indicating that the 5-year survival rate of patients who underwent surgery was extremely increased.

According to the latest cancer statistics ¹, BC has replaced lung cancer as the most common cancer in the America. Consequently, BC treatment and handling the corresponding complications have brought a heavy medical burden to the society, which is a major problem associated with human cancer. In particular, the quality of life and survival rate of BC patients are significantly reduced after they develop BC. This is mainly attributed to the occurrence of skeleton-related events (SREs) after BM. Studies have revealed that the severe vertebral invasion, pathological fractures, bone pain, and other SREs pose a serious threat to the prognosis of patients with BM ^20,21. A previous survey found that the cumulative incidence of SREs in patients with BM is about 45.1% ²². In the present study, the incidence of BM in BC patients was about 2.3%, which reflects the difficulty of diagnosing patients with BM and the harmfulness of BM. Therefore, it is necessary to effectively screen patients who are prone to develop BM and have poor prognosis after BM. Bone scan combined with CT is the gold standard for detecting BM, and is also the preferred method recommended in current guidelines ^23–25. A recent European prospective study showed that [¹⁸F] FDG PET/MRI and MRI were significantly better than CT or bone scintigraphy for the detection of BM in newly diagnosed BC patients ²⁶. However, these tests have some drawbacks, such as radiation damage and high cost, and not all patients are willing to undergo BM testing. Thus, to more effectively address these issues, this study aimed at developing two facile clinical models for early detection of high-risk BCBM patients and prediction of BCBM prognosis.

With the rapid development of artificial intelligence technology, machine learning is increasingly being applied in the field of biomedicine, and it also has great potential in future clinical practice ^27,28. In 2022, an article published in the journal Nature by Stephen-John Sammut et al. presented a study encompassing clinical information, pathology, genomics, and transcriptomics of 168 patients with breast cancer undergoing chemotherapy. They successfully predicted the complete response of chemotherapy patients using a multi-group machine learning approach (AUC = 0.87) ²⁹. This groundbreaking study demonstrates the significant medical value that mature machine learning models can offer in clinical practice, enabling the provision of more accurate assistance to doctors and patients through alternative methods.

Nevertheless, despite significant advancements in building and utilizing various models, there is still considerable scope for improvement. Li Li et al. developed a deep learning algorithm that predicts bone metastasis in breast cancer by incorporating MRI radiological features from 96 cases of metastatic breast tumors and 192 cases of non-metastatic breast tumors. The predictive performance of the model is evaluated using statistical morphology and grayscale characteristics, employing metrics such as AUC, sensitivity, and specificity ³⁰. Nonetheless, due to the high demand for front-end MRI images and data, this model cannot be widely adopted. Thio et al. integrated survival data from thousands of cancer patients with extensive bone metastasis to develop a survival prediction model ³¹. However, the majority of the data utilized for model development and verification originates from laboratory sources, including biochemical data, blood routine data, protease data, and more. While accurately predicting the survival rate, it also imposes more stringent demands on the types of data used. Our machine learning model is specifically designed to predict the occurrence of bone metastasis in breast cancer patients and prognosticate patients with bone metastasis. All the parameters required for the model are derived from routine clinical practice, making them more accessible than specific images or laboratory data. This model can also be utilized by hospitals in remote areas or by junior clinicians to guide the comprehensive treatment planning of breast cancer patients, enabling early intervention to prevent and address potential clinical adverse events. Secondly, it employs multiple strategies, such as preventing overfitting and utilizing shrinkage and column subsampling techniques, to enhance algorithmic generalization and learning speed. The XGBoost algorithm, which has demonstrated high accuracy and ease of use in numerous studies ^32–34, is referenced in this model.

This study used univariable and multivariable logistic regression analyses to screen fifteen independent risk factors, including age, race, sex, grade, T stage, N stage, surgery, radiotherapy, chemotherapy, tumor size, brain metastasis, lung metastasis, liver metastasis, breast subtype, and PR. According to the order of importance of the SHAP diagram, the features that contributed prominently were surgery, N stage, and T stage. Next, univariate and multivariate Cox proportional hazards regression analyses were applied to screen thirteen independent prognostic factors, including age, race, marital status, grade, breast subtype, surgery, radiotherapy, chemotherapy, brain metastases, liver metastases, lung metastases, ER, and PR. All features were also ranked by importance, with results showing that surgery, liver metastases, and lung metastases were the three factors strongly associated with prognosis. However, some features that were considered meaningful in the multivariable logistic regression analysis and multivariate Cox proportional hazards regression analysis had a SHAP value of zero in importance ranking. This may further reflect the superiority of machine learning. Specifically, it can better eliminate unnecessary features unlike traditional linear regression analysis, which has the problem of overfitting. Machine learning enables us to obtain more accurate predictive models by continuously improving operational efficiency and self-improvement.

This study found that BC patients who did not undergo surgery were at high risk of developing BM. Yao et al. ¹⁷ also suggested that surgery was an independent risk factor for BCBM. Despite the hazard of radiation damage, we still recommend bone scans to examine BM in unoperated BC patients. We also found that T stage and N stage were strong predictors of BM. Studies have demonstrated that the increase of T and N stages of malignant tumors indicates the increase of tumor volume, and the expansion of the degree and extent of involvement of adjacent tissues and lymph nodes, which are the manifestations of further development of malignant tumors ^35,36. It is well known that the TNM staging system proposed by the AJCC is a widely used prognostic system ³⁷. However, previous studies have shown that the accuracy of using the TNM staging system alone to predict metastases is not high, and thus researchers often obtain better prediction results through comprehensive analysis of multiple factors ^38,39. Interestingly, surgery was also the most prominent feature with regard to prediction of BCBM prognosis. Although metastatic BC remains an incurable disease, surgery to remove the primary tumor is associated with improved survival in patients with distant metastatic BC at diagnosis. One study reported that patients who underwent primary surgery had significantly longer median survival than those who did not, and primary tumor resection for primary BCBM reduced the risk of death by approximately 40% ⁴⁰. A randomized controlled trial conducted in Turkey found that the 3-year OS was similar in patients with and without primary BC surgery. However, at a median follow-up of 5 years, patients who underwent surgery had a prolonged median OS by approximately nine months ⁴¹. In addition, a trial conducted in India, revealed that the OS of patients with de novo metastatic BC was not improved after surgery for their primary BC ⁴². Scholars in Europe concluded that surgical treatment of the primary tumor in patients with de novo metastatic BC could not benefit majority of them ⁴³. A retrospective study by Gong et al. ⁴⁴ identified surgery as an independent prognostic factor for BCBM, which is consistent with our findings. Therefore, whether the primary tumor of BCBM should be operated is still controversial, which calls for further multicenter prospective studies for verification. Liver and lung metastases play an important role in predicting the prognosis of BCBM. This study found that BCBM patients with liver metastasis or lung metastasis had a poor prognosis, and their 5-year survival rate was lower than that of other types of BCBM patients. We comprehensively considered all meaningful features to predict the prognosis of BCBM and achieved good predictive performance.

The ultimate purpose of building models is to be more convenient for clinical application and help clinicians make decisions. Consequently, based on the XGBoost algorithm, we built two accessible online websites (https://share.streamlit.io/lry4000/bone_metastasis/main) and (https://share.streamlit.io/lry4000/sc5_new/main ). Specifically, a streamlined web page structure enables users to input data more efficiently. The clinical parameters mentioned in the article are displayed on the right side of the webpage, allowing users to input corresponding clinical data based on the actual condition of the patients. The system will instantly generate the predicted probability of bone metastasis for the patient. The results can be presented in various formats and shared with a broader range of clinical participants. The second web page, which predicts the survival rate, follows a similar usage process.

There are some limitations in our study. First, this is a multicenter retrospective study involving only patients from the United States, and thus it inevitably suffers from selection bias. Therefore, there is a need for external data from other countries to validate the reproducibility of our results. Second, although our model achieved good clinical performance on the basis of SEER database, it is essential to further confirm the reliability of the model through prospective studies. Third, the SEER database does not include blood routine, biochemical indicators, and Charlson Comorbidity Index (CCI), which may lead to the model missing some important features.

This study introduced the XGBoost-based machine learning model, for the first time, to construct the diagnosis system and prognosis system for BCBM patients. We sorted the importance of different features using the demographic characteristics and pathological indicators screened from the SEER database. Furthermore, ROC curves, learning curves, precision curves, calibration plots, and decision curves were used to evaluate performance of the model, and an external verification cohort was established to further verify the model. Finally, we have developed two sample and convenient network applications for helping clinicians better achieve clinical decision-making.

ACKNOWLEDGEMENS

We thank all the participants who were involved in this study and the SEER database for providing patient’s data.

AUTHORS’ CONTRIBUTIONS

QB and WZ conceived the study. XZ collected and analyzed the data. XZ and YL drew tables and graphs. WZ and YL jointly prepared this manuscript. All authors contributed to the article and approved the submitted version.

FUNDING

This work was funded by the Key Research and Development Program of Zhejiang Province (2021C03078) and the Health Science and Technology Program of Zhejiang Province (2022KY583).

AVAILABILITY OF DATA AND METARIAL

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

The data for this study were obtained from the SEER database. Patient information in this database is public and anonymous, since ethical approval and informed consent are not required for this study.

CONSENT FOR PUBLICATION

All authors have reviewed the final version of the manuscript and approved it for publication.

DISCLOSURE STATEMENT

The authors declare that they do not have any conflicts of interest.

Siegel, R. L., Miller, K. D., Fuchs, H. E. & Jemal, A. Cancer statistics, 2022. CA: a cancer journal for clinicians 72, 7-33, doi:10.3322/caac.21708 (2022).
DeSantis, C. E., Ma, J., Goding Sauer, A., Newman, L. A. & Jemal, A. Breast cancer statistics, 2017, racial disparity in mortality by state. CA: a cancer journal for clinicians 67, 439-448, doi:10.3322/caac.21412 (2017).
Li, Z. & Kang, Y. Emerging therapeutic targets in metastatic progression: A focus on breast cancer. Pharmacology & therapeutics 161, 79-96, doi:10.1016/j.pharmthera.2016.03.003 (2016).
Schrijver, W. et al. Mutation Profiling of Key Cancer Genes in Primary Breast Cancers and Their Distant Metastases. Cancer research 78, 3112-3121, doi:10.1158/0008-5472.can-17-2310 (2018).
Ng, C. K. Y. et al. Genetic Heterogeneity in Therapy-Naïve Synchronous Primary Breast Cancers and Their Metastases. Clin Cancer Res 23, 4402-4415, doi:10.1158/1078-0432.ccr-16-3115 (2017).
Liang, Y., Zhang, H., Song, X. & Yang, Q. Metastatic heterogeneity of breast cancer: Molecular mechanism and potential therapeutic targets. Seminars in cancer biology 60, 14-27, doi:10.1016/j.semcancer.2019.08.012 (2020).
Tulotta, C. & Ottewell, P. The role of IL-1B in breast cancer bone metastasis. Endocrine-related cancer 25, R421-r434, doi:10.1530/erc-17-0309 (2018).
Jin, L. et al. Breast cancer lung metastasis: Molecular biology and therapeutic implications. Cancer biology & therapy 19, 858-868, doi:10.1080/15384047.2018.1456599 (2018).
Allemani, C. et al. Global surveillance of trends in cancer survival 2000-14 (CONCORD-3): analysis of individual records for 37 513 025 patients diagnosed with one of 18 cancers from 322 population-based registries in 71 countries. Lancet (London, England) 391, 1023-1075, doi:10.1016/s0140-6736(17)33326-3 (2018).
Valastyan, S. & Weinberg, R. A. Tumor metastasis: molecular insights and evolving paradigms. Cell 147, 275-292, doi:10.1016/j.cell.2011.09.024 (2011).
Xiong, Z. et al. Bone metastasis pattern in initial metastatic breast cancer: a population-based study. Cancer management and research 10, 287-295, doi:10.2147/cmar.s155524 (2018).
Coleman, R. E. Metastatic bone disease: clinical features, pathophysiology and treatment strategies. Cancer treatment reviews 27, 165-176, doi:10.1053/ctrv.2000.0210 (2001).
Chen, Y. C., Sosnoski, D. M. & Mastro, A. M. Breast cancer metastasis to the bone: mechanisms of bone loss. Breast cancer research : BCR 12, 215, doi:10.1186/bcr2781 (2010).
Coleman, R. E. Clinical features of metastatic bone disease and risk of skeletal morbidity. Clin Cancer Res 12, 6243s-6249s, doi:10.1158/1078-0432.ccr-06-0931 (2006).
Burstein, H. J. et al. Customizing local and systemic therapies for women with early breast cancer: the St. Gallen International Consensus Guidelines for treatment of early breast cancer 2021. Annals of oncology : official journal of the European Society for Medical Oncology 32, 1216-1235, doi:10.1016/j.annonc.2021.06.023 (2021).
Tan, P. H. et al. The 2019 World Health Organization classification of tumours of the breast. Histopathology 77, 181-185, doi:10.1111/his.14091 (2020).
Yao, Y. B., Zheng, X. E., Luo, X. B. & Wu, A. M. Incidence, prognosis and nomograms of breast cancer with bone metastases at initial diagnosis: a large population-based study. American journal of translational research 13, 10248-10261 (2021).
Liu, D. et al. Breast Subtypes and Prognosis of Breast Cancer Patients With Initial Bone Metastasis: A Population-Based Study. Frontiers in oncology 10, 580112, doi:10.3389/fonc.2020.580112 (2020).
Lee, C. et al. Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database. The Lancet. Digital health 3, e158-e165, doi:10.1016/s2589-7500(20)30314-9 (2021).
Hatoum, H. T., Lin, S. J., Smith, M. R., Barghout, V. & Lipton, A. Zoledronic acid and skeletal complications in patients with solid tumors and bone metastases: analysis of a national medical claims database. Cancer 113, 1438-1445, doi:10.1002/cncr.23775 (2008).
Coleman, R. E. Skeletal complications of malignancy. Cancer 80, 1588-1594, doi:10.1002/(sici)1097-0142(19971015)80:8+<1588::aid-cncr9>3.3.co;2-z (1997).
Hong, S., Youk, T., Lee, S. J., Kim, K. M. & Vajdic, C. M. Bone metastasis and skeletal-related events in patients with solid cancer: A Korean nationwide health insurance database study. PloS one 15, e0234927, doi:10.1371/journal.pone.0234927 (2020).
Cardoso, F. et al. 4th ESO-ESMO International Consensus Guidelines for Advanced Breast Cancer (ABC 4)†. Annals of oncology : official journal of the European Society for Medical Oncology 29, 1634-1657, doi:10.1093/annonc/mdy192 (2018).
Gradishar, W. J. et al. Breast Cancer, Version 4.2017, NCCN Clinical Practice Guidelines in Oncology. Journal of the National Comprehensive Cancer Network : JNCCN 16, 310-320, doi:10.6004/jnccn.2018.0012 (2018).
Cardoso, F. et al. 5th ESO-ESMO international consensus guidelines for advanced breast cancer (ABC 5). Annals of oncology : official journal of the European Society for Medical Oncology 31, 1623-1649, doi:10.1016/j.annonc.2020.09.010 (2020).
Bruckmann, N. M. et al. Prospective comparison of the diagnostic accuracy of 18F-FDG PET/MRI, MRI, CT, and bone scintigraphy for the detection of bone metastases in the initial staging of primary breast cancer patients. European radiology 31, 8714-8724, doi:10.1007/s00330-021-07956-0 (2021).
Goecks, J., Jalili, V., Heiser, L. M. & Gray, J. W. How Machine Learning Will Transform Biomedicine. Cell 181, 92-101, doi:10.1016/j.cell.2020.03.022 (2020).
Hofer, I. S., Burns, M., Kendale, S. & Wanderer, J. P. Realistically Integrating Machine Learning Into Clinical Practice: A Road Map of Opportunities, Challenges, and a Potential Future. Anesthesia and analgesia 130, 1115-1118, doi:10.1213/ane.0000000000004575 (2020).
Sammut, S. J. et al. Multi-omic machine learning predictor of breast cancer therapy response. Nature 601, 623-629, doi:10.1038/s41586-021-04278-5 (2022).
Li, L., Tian, H., Zhang, B., Wang, W. & Li, B. Prediction for Distant Metastasis of Breast Cancer Using Dynamic Contrast-Enhanced Magnetic Resonance Imaging Images under Deep Learning. Comput Intell Neurosci 2022, 6126061, doi:10.1155/2022/6126061 (2022).
Thio, Q. et al. Development and Internal Validation of Machine Learning Algorithms for Preoperative Survival Prediction of Extremity Metastatic Disease. Clin Orthop Relat Res 478, 322-333, doi:10.1097/CORR.0000000000000997 (2020).
Bolourani, S. et al. A Machine Learning Prediction Model of Respiratory Failure Within 48 Hours of Patient Admission for COVID-19: Model Development and Validation. Journal of medical Internet research 23, e24246, doi:10.2196/24246 (2021).
Hou, N. et al. Predicting 30-days mortality for MIMIC-III patients with sepsis-3: a machine learning approach using XGboost. Journal of translational medicine 18, 462, doi:10.1186/s12967-020-02620-5 (2020).
Guan, X. et al. Clinical and inflammatory features based machine learning model for fatal risk prediction of hospitalized COVID-19 patients: results from a retrospective cohort study. Annals of medicine 53, 257-266, doi:10.1080/07853890.2020.1868564 (2021).
Giuliano, A. E. et al. Breast Cancer-Major changes in the American Joint Committee on Cancer eighth edition cancer staging manual. CA: a cancer journal for clinicians 67, 290-303, doi:10.3322/caac.21393 (2017).
Ryzhov, A. et al. Comparison of breast cancer and cervical cancer stage distributions in ten newly independent states of the former Soviet Union: a population-based study. The Lancet. Oncology 22, 361-369, doi:10.1016/s1470-2045(20)30674-4 (2021).
Hortobagyi, G. N., Edge, S. B. & Giuliano, A. New and Important Changes in the TNM Staging System for Breast Cancer. American Society of Clinical Oncology educational book. American Society of Clinical Oncology. Annual Meeting 38, 457-467, doi:10.1200/edbk_201313 (2018).
Hu, C. et al. Diagnostic and prognostic nomograms for bone metastasis in hepatocellular carcinoma. BMC cancer 20, 494, doi:10.1186/s12885-020-06995-y (2020).
Chen, B. et al. Risk Factors, Prognostic Factors, and Nomograms for Distant Metastasis in Patients With Newly Diagnosed Osteosarcoma: A Population-Based Study. Frontiers in endocrinology 12, 672024, doi:10.3389/fendo.2021.672024 (2021).
Ruiterkamp, J. et al. Surgical resection of the primary tumour is associated with improved survival in patients with distant metastatic breast cancer at diagnosis. European journal of surgical oncology : the journal of the European Society of Surgical Oncology and the British Association of Surgical Oncology 35, 1146-1151, doi:10.1016/j.ejso.2009.03.012 (2009).
Soran, A. et al. Randomized Trial Comparing Resection of Primary Tumor with No Surgery in Stage IV Breast Cancer at Presentation: Protocol MF07-01. Annals of surgical oncology 25, 3141-3149, doi:10.1245/s10434-018-6494-6 (2018).
Badwe, R. et al. Locoregional treatment versus no treatment of the primary tumour in metastatic breast cancer: an open-label randomised controlled trial. The Lancet. Oncology 16, 1380-1388, doi:10.1016/s1470-2045(15)00135-7 (2015).
Poggio, F., Lambertini, M. & de Azambuja, E. Surgery of the primary tumour in patients presenting with de novo metastatic breast cancer: to do or not to do? ESMO open 3, e000324, doi:10.1136/esmoopen-2018-000324 (2018).
Gong, Y. et al. Incidence proportions and prognosis of breast cancer patients with bone metastases at initial diagnosis. Cancer medicine 7, 4156-4169, doi:10.1002/cam4.1668 (2018).

Table 1 Demographic and clinicopathological characteristics in breast cancer patients.

Characteristic

Training cohort(n=198364)

No. of patients %

Validation cohort(n=85009)

No. of patients %

P value

Age

20-39y

40-59y

60-79y

≥80y

Race

Black

White

Other

Sex

Female

Male

Laterality

Left

Right

Marital status

Married

Single

Divorced

Widowed

Separated

Unmarried or Domestic Partner

Grade

Well differentiated: I

Moderately differentiated: II

Poorly differentiated: III

Undifferentiated; anaplastic: IV

AJCC T stage

T1

T2

T3

T4

AJCC N stage

N0

N1

N2

N3

Surgery

No

Yes

Radiotherapy

No

Yes

Chemotherapy

No

Yes

Tumor size

Median

Range

Bone metastasis

No

Yes

Brain metastasis

No

Yes

Lung metastasis

No

Yes

Liver metastasis

No

Yes

Breast subtype

HR+/HER2- (Luminal A)

HR+/HER2+ (Luminal B)

HR-/HER2+ (HER2 enriched)

HR-/HER2- (Triple Negative)

ER

Negative

Positive

PR

Negative

Positive

HER2

Negative

Positive

10782

83651

87370

16561

21976

156913

19475

196954

1410

100382

97982

116314

30877

22170

26048

2257

698

45730

88147

63997

490

116300

62675

12697

6692

134065

47260

10669

6370

10662

187702

84483

113881

114342

84022

18

11-28

193773

4591

197953

411

196215

2149

196593

1771

146405

21392

8626

21941

32525

165839

52466

145898

168346

30018

5.4

42.2

44.1

8.3

11.1

79.1

9.8

99.3

0.7

50.6

49.4

58.6

15.6

11.2

13.1

1.1

0.4

23.1

44.4

32.3

0.2

58.6

31.6

6.4

3.4

67.6

23.8

5.4

3.2

5.4

94.6

42.6

57.4

57.6

42.4

97.7

2.3

99.8

0.2

98.9

1.1

99.1

0.9

73.8

10.8

4.3

11.1

16.4

83.6

26.4

73.6

84.9

15.1

4609

35853

37592

6955

9421

67488

8100

84406

603

43143

41866

49811

13210

9678

11030

970

310

19328

37509

27965

207

49704

27060

5391

2854

57517

20212

4533

2747

4574

80435

36119

48890

48551

36458

18

11-28

83154

1855

84828

181

84151

858

84280

729

62383

9217

3816

9593

14264

70745

22862

62147

71976

13033

5.4

42.2

44.2

8.2

11.1

79.4

9.5

99.3

0.7

50.8

49.2

58.6

15.5

11.4

13.0

1.1

0.4

22.7

44.1

32.9

0.2

58.5

31.8

6.3

3.4

67.7

23.8

5.3

3.2

5.4

94.6

42.5

57.5

57.1

42.9

97.8

2.2

99.8

0.2

99.0

1.0

99.1

0.9

73.4

10.8

4.5

11.3

16.8

83.2

26.9

73.1

84.7

15.3

0.485

0.057

0.966

0.476

0.589

0.010

0.641

0.936

0.951

0.617

0.009

0.840#

0.030

0.760

0.078

0.358

0.073

0.012

0.014

0.177

#Mann-Whitney U test

Table 2 Demographic and clinicopathological features in breast cancer patients with bone metastasis.

Characteristic

Training cohort(n=2448)

No. of patients %

Validation cohort(n=1044)

No. of patients %

P value

Age

20-39y

40-59y

60-79y

≥80y

Race

Black

White

Other

Sex

Female

Male

Laterality

Left

Right

Marital status

Married

Single

Divorced

Widowed

Separated

Unmarried or Domestic Partner

Grade

Well differentiated: I

Moderately differentiated: II

Poorly differentiated: III

Undifferentiated; anaplastic: IV

AJCC T stage

T1

T2

T3

T4

AJCC N stage

N0

N1

N2

N3

Surgery

No

Yes

Radiotherapy

No

Yes

Chemotherapy

No

Yes

Tumor size

Median

Range

Brain metastasis

No

Yes

Lung metastasis

No

Yes

Liver metastasis

No

Yes

Breast subtype

HR+/HER2- (Luminal A)

HR+/HER2+ (Luminal B)

HR-/HER2+ (HER2 enriched)

HR-/HER2- (Triple Negative)

ER

Negative

Positive

PR

Negative

Positive

HER2

Negative

Positive

174

985

1007

282

440

1839

169

2416

32

1220

1228

1099

564

343

395

39

8

196

1068

1170

14

303

832

491

822

561

1144

325

418

1606

842

1386

1062

1138

1310

44

27-70

2264

184

1752

696

1789

659

1588

358

345

157

537

1911

900

1548

1933

515

7.1

40.2

41.1

11.5

18.0

75.1

6.9

98.7

1.3

49.8

50.2

44.9

23.0

14.0

16.1

1.6

0.3

8.0

43.6

47.8

0.6

12.4

34.0

20.1

33.6

22.9

46.7

13.3

17.1

65.6

34.4

56.6

43.4

46.5

53.5

92.5

7.5

71.6

28.4

73.1

26.9

64.9

14.6

14.1

6.4

21.9

78.1

36.8

63.2

79.0

21.0

76

402

452

114

186

787

71

1025

19

543

501

452

259

125

185

21

2

78

481

480

5

129

402

199

314

256

491

139

158

688

356

596

448

515

529

40

27-62

950

94

737

307

776

268

707

147

132

58

201

843

355

689

839

205

7.3

38.5

43.3

10.9

17.8

75.4

6.8

98.2

1.8

52.0

48.0

43.3

24.8

12.0

17.7

2.0

0.2

7.5

46.1

46.0

0.4

12.4

38.5

19.1

30.1

24.5

47.0

13.3

15.1

65.9

34.1

57.1

42.9

49.3

50.7

91.0

9.0

70.6

29.4

74.3

25.7

67.7

14.1

12.6

5.6

19.3

80.7

34.0

66.0

80.4

19.6

0.652

0.986

0.248

0.239

0.304

0.597

0.061

0.477

0.866

0.797

0.124

0.034#

0.137

0.560

0.444

0.383

0.075

0.120

0.349

#Mann-Whitney U test

Table 3 Univariable and multivariable Logistic regression of risk factors of bone metastasis in breast cancer patients.

Characteristics

Univariate analysis

Multivariate analysis

OR (95%CI)

P

OR (95%CI)

P

Age

20-39y

40-59y

60-79y

≥80y

Race

Black

White

Other

Sex

Female

Male

Laterality

Left

Right

Marital status

Married

Single

Divorced

Widowed

Separated

Unmarried or Domestic Partner

Grade

Well differentiated : I

Moderately differentiated : II

Poorly differentiated : III

Undifferentiated; anaplastic :IV

AJCC T stage

T1

T2

T3

T4

AJCC N stage

N0

N1

N2

N3

Surgery

No

Yes

Radiotherapy

No

Yes

Chemotherapy

No

Yes

Tumor size

Brain metastasis

No

Yes

Lung metastasis

No

Yes

Liver metastasis

No

Yes

Breast subtype

HR+/HER2- (Luminal A)

HR+/HER2+ (Luminal B)

HR-/HER2+ (HER2 enriched)

HR-/HER2- (Triple Negative)

ER

Negative

Positive

PR

Negative

Positive

HER2

Negative

Positive

Reference

0.661(0.591-0.741)

0.616(0.551-0.691)

0.705(0.611-0.813)

Reference

0.694(0.640-0.754)

0.597(0.525-0.677)

Reference

2.291(1.789-2.887)

Reference

0.950(0.896-1.007)

Reference

1.871(1.737-2.013)

1.342(1.221-1.473)

1.302(1.190-1.422)

1.746(1.368-2.193)

1.131(0.647-1.820)

Reference

2.750(2.479-3.058)

3.414(3.074-3.800)

3.877(2.279-6.150)

Reference

5.654(5.142-6.225)

15.907(14.310-17.698)

54.099(48.922-59.901)

Reference

6.404(5.942-6.906)

8.215(7.418-9.089)

17.608(15.960-19.417)

Reference

0.023(0.021-0.024)

Reference

0.534(0.503-0.567)

Reference

1.907(1.797-2.023)

1.027(1.026-1.027)

Reference

86.763(70.737-106.884)

Reference

64.351(58.746-70.505)

Reference

80.236(72.580-88.748)

Reference

1.848(1.707-1.998)

1.644(1.451-1.853)

0.942(0.849-1.041)

Reference

0.977(0.904-1.057)

Reference

0.816(0.765-0.869)

Reference

1.802(1.682-1.930)

<0.001

0.085

<0.001

0.638

<0.001

0.251

0.566

<0.001

Reference

0.905(0.786-1.045)

0.953(0.823-1.106)

0.527(0.431-0.645)

Reference

1.195(1.071-1.336)

0.822(0.697-0.967)

Reference

2.026(1.493-2.702)

Reference

1.336(1.182-1.513)

1.091(0.954-1.249)

0.733(0.361-1.384)

Reference

2.751(2.461-3.079)

3.843(3.307-4.468)

6.137(5.294-7.119)

Reference

2.635(2.399-2.895)

4.161(3.645-4.745)

6.648(5.810-7.603)

Reference

0.048(0.044-0.053)

Reference

1.410(1.296-1.534)

Reference

0.762(0.698-0.836)

1.002(1.000-1.003)

Reference

4.605(3.439-6.188)

Reference

4.718(4.156-5.357)

Reference

9.341(8.141-10.723)

Reference

0.880(0.789-0.981)

0.423(0.358-0.499)

0.436(0.358-0.529)

Reference

1.121(1.004-1.253)

0.168

0.523

<0.001

0.002

0.019

<0.001

0.205

0.363

<0.001

0.022

<0.001

0.043

Table 4 Univariate and multivariate cox regression of prognostic factors of bone metastasis in breast cancer patients.

Characteristics

Univariate analysis

Multivariate analysis

HR (95%CI)

P

HR (95%CI)

P

Age

20-39y

40-59y

60-79y

≥80y

Race

Black

White

Other

Sex

Female

Male

Laterality

Left

Right

Marital status

Married

Single

Divorced

Widowed

Separated

Unmarried or Domestic Partner

Grade

Well differentiated : I

Moderately differentiated : II

Poorly differentiated : III

Undifferentiated; anaplastic :IV

AJCC T stage

T1

T2

T3

T4

AJCC N stage

N0

N1

N2

N3

Surgery

No

Yes

Radiotherapy

No

Yes

Chemotherapy

No

Yes

Tumor size

Brain metastasis

No

Yes

Lung metastasis

No

Yes

Liver metastasis

No

Yes

Breast subtype

HR+/HER2- (Luminal A)

HR+/HER2+ (Luminal B)

HR-/HER2+ (HER2 enriched)

HR-/HER2- (Triple Negative)

ER

Negative

Positive

PR

Negative

Positive

HER2

Negative

Positive

Reference

1.096(0.922-1.304)

1.409(1.185-1.674)

1.993(1.632-2.434)

Reference

0.731(0.656-0.814)

0.727(0.604-0.874)

Reference

1.019(0.715-1.453)

Reference

1.034(0.951-1.124)

Reference

1.259(1.131-1.401)

1.289(1.135-1.467)

1.579(1.401-1.779)

1.233(0.884-1.719)

2.109(1.051-4.232)

Reference

1.261(1.067-1.491)

1.572(1.332-1.856)

2.131(1.233-3.683)

Reference

1.051(0.912-1.212)

1.187(1.019-1.384)

1.370(1.189-1.578)

Reference

1.015(0.912-1.130)

0.886(0.766-1.024)

0.944(0.826-1.078)

Reference

0.500(0.457-0.548)

Reference

0.735(0.675-0.799)

Reference

0.797(0.733-0.867)

1.003(1.002-1.004)

Reference

2.124(1.824-2.474)

Reference

1.648(1.504-1.806)

Reference

1.848(1.683-2.028)

Reference

1.135(1.006-1.281)

2.141(1.895-2.418)

1.274(1.072-1.514)

Reference

0.565(0.511-0.624)

Reference

0.619(0.568-0.676)

Reference

1.055(0.953-1.169)

0.300

<0.001

0.915

0.432

<0.001

0.218

0.036

0.007

<0.001

0.007

0.489

0.028

<0.001

0.784

0.099

0.394

<0.001

0.039

<0.001

0.006

<0.001

0.304

Reference

1.074(0.899-1.283)

1.389(1.158-1.669)

1.831(1.459-2.298)

Reference

0.784(0.700-0.879)

0.849(0.703-1.028)

Reference

1.158(1.035-1.296)

1.128(0.990-1.286)

1.185(1.038-1.354)

1.103(0.787-1.544)

2.075(1.028-4.189)

Reference

1.258(1.062-1.489)

1.439(1.209-1.713)

1.437(0.826-2.500)

Reference

0.569(0.517-0.628)

Reference

0.809(0.741-0.884)

Reference

0.713(0.645-0.789)

Reference

1.819(1.551-2.135)

Reference

1.286(1.168-1.415)

Reference

1.661(1.499-1.839)

Reference

0.979(0.856-1.122)

0.831(0.563-1.226)

0.429(0.285-0.646)

Reference

0.482(0.339-0.686)

Reference

0.702(0.621-0.793)

0.428

<0.001

0.094

0.011

0.070

0.012

0.569

0.042

0.008

<0.001

0.199

<0.001

0.769

0.352

<0.001

Table 5 The performance of diagnostic models on six ML algorithms.
Models	Training cohort					Validation cohort
Models	AUC	Accuracy	Precision	Recall	F1-score	AUC	Accuracy	Precision	Recall	F1-score
XGBoost RF DT ET GaussianNB LR	0.987 0.963 0.928 0.918 0.810 0.859	0.947 0.888 0.850 0.821 0.669 0.769	0.948 0.889 0.868 0.839 0.718 0.775	0.947 0.888 0.850 0.821 0.669 0.769	0.947 0.888 0.849 0.819 0.649 0.768	0.939 0.932 0.894 0.894 0.813 0.861	0.938 0.846 0.953 0.917 0.894 0.797	0.978 0.976 0.976 0.975 0.967 0.973	0.938 0.846 0.953 0.917 0.894 0.797	0.954 0.899 0.962 0.941 0.926 0.868

Abbereviations: XGBoost, Extreme gradient boosting; RF, Random Forest; DT, Decision Tree; ET, Extra Tree; GaussianNB, Gaussian Naive Bayesian; LR, Logistic Regression

Supplementary Figures S1-S2 are not available with this version.

No competing interests reported.

Predicting diagnosis and survival of bone metastasis in breast cancer using machine learning

Status:

Journal Publication

Version 1

Abstract

Figures

INTRODUCTION

MATERIALS AND METHODS

Study population

Feature selection and validation strategy

Machine learning algorithms

Statistical analysis

RESULTS

DISCUSSION

CONCLUSION

Declarations

References

Tables

Supplementary Figures

Additional Declarations

Status:

Journal Publication

Version 1