The GLOBOCAN 2020 report highlights breast cancer (BC) as the global cancer with the highest incidence worldwide (Rahimzadeh et al., 2021). This report is underscored by its alarming mortality rates, accounting for 685,000 deaths in 2020 and establishing it as the foremost cause of cancer-related mortality in women. Representing 11% of all cancer cases, breast cancer diagnoses in women totaled approximately 2.3 million. A notable escalation of 20% in prevalence and a 14% increase in global mortality rates have been observed since the 2008 report (Kashyap et al., 2022).

Projections from global breast cancer surveillance studies indicate a looming 40% rise in new cases, surpassing three million, with a simultaneous 50% surge in breast cancer-related deaths, reaching one million by 2040 (Arnold et al., 2022). In addition, BC increases female deaths significantly in underdeveloped countries, ranks fifth overall in cancer-related deaths, and is the primary cause of death among women in these regions. In developed countries, breast cancer ranks second after lung cancer in cancer-related deaths (Cao et al., 2021).

While the average 5-year survival rate for BC at all stages hovers around 90% (Siegel et al., 2018), the 5-year survival rate significantly rises to nearly 99% for early-stage BC (comprising stages I and II). Unfortunately, only 61% of patients receive an early-stage diagnosis (DeSantis et al., 2017). Despite the demonstrated high sensitivity (>90%) of various clinical and experimental imaging techniques in diagnosing symptomatic BC patients, there remains a conspicuous absence of an effective diagnostic tool for early-stage BC (EBC) detection (Fiorica, 2016; Bonilla et al., 2017). Consequently, there is an urgent demand for a detection method that ensures precise and early diagnosis and screening, ultimately facilitating prompt treatment and enhancing overall BC survival rates.

Examining metabolites within biological pathways offers crucial insights into the functional phenotype of tumors. The exploration of metabolic alterations in cancer cells dates back to the first half of the 20th century, notably with Otto Warburg’s discovery of aerobic glycolysis, now recognized as the Warburg effect, a prevalent feature in various tumors (Tenori et al., 2015; Schwartz et al., 2017). Furthermore, metabolites contribute essential information that intersects with fields such as proteomics, transcriptomics, and genomics, enhancing our systematic understanding of biological systems.

Although numerous studies have shown promise in investigating the metabolomic profile of BC, these findings require additional validation before integration into clinical practice (Tenori et al., 2015; Jobard et al., 2014; Tenori et al., 2012; Oakman et al., 2011; Wei et al., 2013; Hart et al., 2017) Moreover, existing research lacks an exploration of metabolomic biomarkers for BC through the lens of explainable artificial intelligence (XAI).

Given the potential of the metabolomic approach and the imperative for more comprehensive metabolomic investigations in BC, our study seeks to identify potential biomarkers for early BC detection. Additionally, we aim to develop an interpretable prediction model using XAI to enhance the understanding of metabolomic biomarkers in the context of BC.

This research utilized publicly available (open-access) targeted metabolomics panel data to differentiate BC patients and healthy controls (Jasbi et al., 2019). The study was conducted according to the principles of the Declaration of Helsinki and was approved by the Inonu University Health Sciences Non-Interventional Clinical Research Ethics Committee (protocol code = 2024/5750). A total of 201 samples were included in the study, collected from individuals who underwent an overnight fast; among them, 102 were BC patients and 99 were healthy controls. Controls were matched with BC patients based on age. The study aimed to identify potential metabolic biomarkers through feature selection and explainable artificial intelligence techniques, in addition to the development of the ML prediction model that could increase sensitivity and specificity in detecting the early stage of BC. A targeted liquid chromatography-tandem mass spectrometry (LC-MS/MS) approach was used to examine the metabolomic panel. Further information regarding sample preparation and the conditions of liquid chromatography and mass spectrometry can be found in the Supplementary Material (Jasbi et al., 2019).

The RF algorithm is a popular ensemble learning technique extensively employed in machine learning. It operates by amalgamating numerous decision trees to construct a robust predictive model. In an RF model, each decision tree is developed by training on a subset of data, where different features are assessed against varying examples. Subsequently, every tree generates its individual predictions, and the collective prediction is determined based on the average of these forecasts or the majority vote among the classes. RF exhibits resilience against overfitting and tends to exhibit superior performance particularly on extensive datasets. Additionally, its utility extends to discerning the significance of variables owing to its capability to gauge the relative importance levels among features (Breiman, 2001; Yagin et al., 2024).

GB stands as a robust ML algorithm employed in constructing predictive models. It achieves this by amalgamating various decision trees, culminating in a potent predictive model. Each tree undergoes sequential training, with a focus on rectifying errors that its predecessor failed to address. Leveraging this sequential structure, GB iteratively integrates new trees into training with the aim of error minimization. The efficacy of GB is contingent upon meticulous parameter adjustments; otherwise, it may lead to overfitting. Hence, its application warrants careful consideration to attain a well-balanced model (Yagin B. et al., 2023; Yagin F. H. et al., 2023).

LR, despite its name implying regression, serves as a ML algorithm predominantly employed in classification tasks. Its primary function is to discriminate between two or more classes, typically presenting its output in the form of a probability value. Operating on a linear equation, LR models the correlation between input features and output classes, with the equation’s outcome transformed into a classification decision through a logistic function. Throughout the training process, it employs optimization techniques such as the maximum likelihood method or gradient descent to ascertain the optimal parameters. Despite its advantages in simplicity, interpretability, and swiftness, the efficacy of LR might diminish in scenarios where it fails to discern linear boundaries, encountering difficulty in capturing intricate relationships (Khandezamin et al., 2020; Donisi et al., 2022; Dedeturk and Akay, 2020).

SVMs represent a potent ML algorithm extensively applied in classification tasks. SVMs aim to delineate a decision boundary between two classes and position data points optimally within these bounds. The hallmark of SVMs lies in their ability to generalize beyond the dataset, ensuring data points are segregated by the widest possible margin along the decision boundary. This algorithm operates by leveraging support vectors, which effectively facilitate the clearest separation between distinct classes. Moreover, SVMs exhibit efficacy in handling multidimensional datasets and possess the capability to segregate datasets that lack linear separability by mapping them onto high-er-dimensional spaces. Nevertheless, while dealing with sizable datasets, SVMs encounter heightened computational complexity, demanding precise parameter settings (Sheykhmousa et al., 2020; Guo et al., 2021; Li and Sun, 2020).

nestedCV is a cross-validation technique employed to robustly assess the performance of a model. This approach integrates a two-tiered cross-validation framework. The outer loop segregates the dataset into training and testing folds, while the inner cross-validation structure is applied within each training fold. Within the inner loop, each training fold is further divided into smaller folds, utilized for hyper-parameter tuning or selection of the model. Consequently, a model trained with optimal hyper-parameters is derived for each test fold in the outer loop, and an aggregated performance metric is computed by consolidating performance evaluations. This method finds extensive application in reliably conducting both hyper-parameter optimization and performance evaluations of the model (Wainer and Cawley, 2021; Yates et al., 2023; Hosseini et al., 2020).

RS hyper-parameter optimization is a technique utilized to enhance the performance of ML models by optimizing hyper-parameters. This method involves a random selection of values across hyper-parameters impacting the model’s performance. These selections are made randomly throughout the potential value distributions within a given hyper-parameter space. Through repetitive training and testing iterations using these randomly chosen hyper-parameter values, the objective is to maximize model performance or minimize error rates. RS often proves more efficient than Grid Search as it achieves favorable outcomes by making fewer attempts via random selections, avoiding exhaustive exploration of the entire hyper-parameter space. Particularly effective with expansive hyper-parameter domains, it reduces time investment, offering a more cost-effective approach. Widely favored for its adaptability and time-saving attributes, this method is widely preferred in machine learning applications (Bergstra and Bengio, 2012; Villalobos-Arias and Quesada-López, 2021; Bhat et al., 2018).

SHAP stands as an explainability technique utilized to elucidate the predictions of ML models and comprehend feature contributions. SHAP leverages Shapley values to assess the impact of individual features on predictions, aggregating these contributions to form the model’s output. Inspired by coalition games, Shapley values are employed to quantify each feature’s influence on predictions, considering their interactions with other features. This method furnishes an explanation value for each feature, aiding in discerning the significance of features in understanding the model’s predictions. Noteworthy is SHAP’s effectiveness in handling large datasets and complex models, rendering the model’s inner workings more comprehensible. Despite the additional computational costs associated with SHAP, its interpretability advantages and compliance with XAI principles make it a robust option for biomarker discovery in this study. By leveraging SHAP’s ability to provide a detailed breakdown of feature effects, we aim to increase the transparency, reliability, and clinical relevance of our predictive model for breast cancer diagnosis (Lundberg et al., 2020; Bifarin, 2023; Zheng et al., 2023).

In the initial phase of the proposed approach, the objective is to conduct hyper-parameter optimization and compute SHAP values utilizing the nestedCV method coupled with the random search method for EBC diagnosis. Subsequently, the second stage involves employing the SHAP values obtained in the first stage for feature selection, followed by repeating the training procedures from the initial phase. The flowchart illustrating the proposed approach is depicted in Figure 1.

Figure 1. Flowchart diagram of the proposed approach generated using nestedCV, random search and SHAP computation.

As shown in Figure 1, the proposed model employs two different segmentation strategies: Inner CV and Outer CV. The Inner CV is used to perform hyper-parameter optimization and select the most suitable model. In this step, a 5-fold cross-validation is applied to each training fold of the Outer CV. On the other hand, the Outer CV is designed to assess the overall performance of the model and estimate its generalization error. For this purpose, a 10-fold cross-validation is performed on the entire dataset. Figure 1 demonstrates that within the proposed system, an initial step involves internal cross-validation to conduct hyper-parameter optimization through random search. Following this phase, model training proceeds using Outer CV, accompanied by the computation of SHAP values for the models. Subsequently, model performance metrics are computed based on the specified criteria. Should the model training processes utilize the complete set of features, feature selection is then implemented employing SHAP values, resulting in the creation of a new dataset. The finalization of the proposed approach entails reiteration of hyper-parameter optimization and training processes utilizing the newly generated dataset. Subsequently, the most significant features identified using the SHAP method were analyzed in the context of breast cancer diagnosis. Following this interpretative analysis, candidate features were selected for biomarker development, culminating in the final design of the biomarker.

Table 1 presents the results regarding the patients’ clinical data. The average age of patients in the BC group was higher at 54.637 ± 10.384 years compared to 51.556 ± 12.334 years in the control group. Among the BC patients, the distribution of menopausal status was as follows: 42.16% (43/102) were premenopausal, 5.88% (6/102) were perimenopausal, and the majority, 51.96% (53/102), were postmenopausal. Tumor laterality was predominantly observed in the right breast (53.92%, 55/102), followed by the left breast (42.16%, 43/102), with 3.92% (4/102) of cases involving bilateral tumors. The distribution of cancer stages showed that 23.53% (24/102) of patients were diagnosed at Stage I, 41.18% (42/102) at Stage II, and 35.29% (36/102) at Stage III. Regarding hormone receptor status, 93.14% (95/102) of patients were estrogen receptor (ER) positive, while only 6.86% (7/102) were ER negative. Progesterone receptor (PR) positivity was observed in 86.27% (88/102) of patients, with 13.73% (14/102) being PR negative. HER2 status was negative in 67.65% (69/102) of patients, while 12.75% (13/102) were borderline/intermediate, and 19.61% (20/102) were HER2 positive. Subtype distribution revealed that the majority of patients were ER/PR + HER2- (63.73%, 65/102), followed by ER/PR + HER2+ (17.65%, 18/102), ER/PR- HER2- (3.92%, 4/102), and ER/PR- HER2+ (1.96%, 2/102). The subtype was unknown in 12.75% (13/102) of cases. Furthermore, ductal carcinoma was the most common histological type, present in 90.20% (92/102) of cases, while lobular carcinoma and mixed ductal/lobular carcinoma accounted for 7.84% (8/102) and 1.96% (2/102) (Table 1).

Table 1. Data analysis on patients’ clinical information.

In our study, the proposed approach was implemented using a dataset comprising 201 clinical records, encompassing 102 BC patients and 99 control. The initial experimental phase involved developing the classification and hyper-parameter optimization stages of the proposed model utilizing Python’s sklearn library, specifically leveraging the RandomizedSearchCV, StratifiedKFold, and cross_val_score classes. StratifiedKFold was employed to ensure a balanced representation of samples across various classes within the folds. For inner cross-validation, a 5-fold setup was adopted, whereas 10 folds were utilized for outer cross-validation. The RandomizedSearchCV library facilitates random selections within the provided hyper-parameter space and conducts a set number of trials to identify the hyper-parameter configuration that yields optimal results. Within the proposed approach, the RandomizedSearchCV library was called 20 times. Adhering to the nestedCV paradigm, for each externally created fold, an internal cross-validation was executed with five folds to attain optimal values. The classification phase of the proposed model encompassed the utilization of four distinct classification methods: RF, GB, LR, and SVM. The classification models were developed using sklearn library of Python language. Table 2 delineates the optimized hyper-parameters for each method, the respective hyper-parameter spaces, and the resultant optimal values obtained for each fold.

Table 2. Hyper-parameter space information and optimum hyper-parameter values for the pro-posed models.

The function represented in Table 2 returned values within specified ranges by taking the starting, ending, and increment values of the specified range function, respectively. The “i” value employed in this function denoted each element within the list generated by the range function. Column “Optimum Hyper-Parameters for Each Fold” in Table 2 showcased the optimal values obtained for each external cross-validation, respectively.

In the subsequent phase of the experiment, the performance metrics of the proposed classification methods were computed utilizing the acquired optimal hyper-parameters. In this context, accuracy, precision, recall, and specificity values for each fold were computed for each method. To gauge the model’s robustness, the standard deviation (std) of the performance values across folds was also determined. Table 3 shows the interfold std values for each metric and the average of the metric scores obtained in each fold.

Table 3. Performance values of proposed models calculated on the testing dataset.

According to the findings presented in Table 3, RF, GB, and SVM exhibited the highest accuracy. However, when considering robustness based on the standard deviation of accuracy across each fold, LR emerged as the most robust method. Among the three high-scoring methods, GB was deemed the most robust. Regarding precision scores, RF not only attained the highest metric score but also demonstrated the most resilience. SVM emerged as the most successful method based on recall value, whereas LR displayed the greatest robustness. In terms of specificity, LR showcased both the highest performance and robustness. Upon comprehensive evaluation, all classification algorithms showed their weakest metric performance in specificity, with LR consistently displaying the most robustness. However, an overall assessment indicates no significant superiority of any classification method over the others.

For the classification models, subsequent to the measurement of metric values, SHAP values were computed utilizing the SHAP Library in Python to gauge the feature impact during the training of each classification model. The importance of metabolite features, as per the obtained SHAP values, is separately listed for each classification. Figure 2 illustrates the importance of the top 20 metabolite features calculated using the SHAP method for each classification model.

Figure 2. Feature importance of classification models calculated using SHAP values.

As depicted in Figure 2, the metabolite importance order varies based on the classification method employed. Notably, according to the SHAP values calculated using trained RF, SVM, and GB models, the feature “glutamic acid” holds the utmost importance for subjectivity. In the case of the trained LR model, “glutamic acid” ranks as the fifth most critical metabolite for BC. Additionally, across all classifications, “glutamic acid,” “hypo-xanthine,” and “ornithine” consistently rank among the top 20 metabolites. Leveraging the data from Figure 2, metabolite biomarker selection was conducted individually, resulting in the identification of the top 20 metabolites for each trained model. The selected metabolite names are detailed in Table 4.

Table 4. Metabolite names that selected using SHAP values computed by trained classification models.

Following the metabolite feature selection based on SHAP values derived separately from each trained RF, LR, GB, and SVM model, the classification models underwent retraining, and subsequent performance scores were computed. At this stage, rather than employing traditional feature selection methods, we utilized the explanations provided by the trained models to identify the most influential features. In this context, SHAP values were calculated separately for each trained model, and the features with the highest SHAP values were selected for inclusion in the new model. Thus, the SHAP method served a dual purpose in this study, facilitating both feature selection and model interpretability. Similar to the initial stage, the nestedCV method was employed, with five inner folds and 10 outer folds. Model hyper-parameters were optimized utilizing spaces in the models, encompassing all metabolites. Table 5 shows the interfold std values for each metric and the average of the metric scores obtained in each fold for models trained using selected metabolite features.

Table 5. Performance values of proposed models calculated on the testing dataset.

Upon examining the results presented in Table 4, it is evident that the LR model, trained with selected metabolite features derived from the SHAP values calculated from the trained SVM model, outperforms other models across all metrics, excluding precision. Notably, this model also exhibits the lowest std values for these metrics. Comparatively, when contrasted with models utilizing all metabolites, the selection of metabolites using SHAP values appears to enhance model performance. To explore potential changes in metabolite importance when only the 20 selected metabolite features were employed, SHAP values were recalculated in the newly trained models, elucidating the models. Consequently, models trained in the initial stage, utilizing all metabolites, were also explained utilizing the SHAP method, doubling as the feature selection method. Figure 3 illustrates metabolite importance graphs for each classification utilized as a feature se-lection method in the first stage, alongside the model exhibiting the highest accuracy in the second stage. Specifically, the importance graphs for the RF, SVM, RF, and LR models trained in the second stage were sequentially drawn for the RF, LR, GB, and SVM models, respectively, trained using all metabolite features in the first stage.

Figure 3. Metabolite importance of classification models calculated using SHAP values after feature selection.

In Figure 3, following the feature selection, it is evident that the importance of metabolite features alters within the trained models. Notably, despite this variation, “glutamic acid” retained its status as the most crucial metabolites across both models. Moreover, it emerged as the second most importance metabolite in the RF model and ranked sixth in importance in LR. Among the other two metabolites common to all four models, “hypoxanthine” consistently ranked ninth at worst in all models, while “ornithine” held the seventh position in the least favourable scenario across all models. Detailed examination of the model explanations using SHAP is crucial. Therefore, SHAP waterfall plots for a selected positive example are presented separately for each model in Figure 4.

Figure 4. Waterfall diagram of classification models calculated using SHAP values after feature selection.

Upon examining these graphs, it is evident that the “glutamic acid” feature holds significant importance across all models. This finding aligns with the previously obtained summary SHAP plots, reinforcing the assumption that this feature plays a crucial role in model training.

In our study, to further investigate the explanations provided by SHAP, we examined the correlation between the selected features. For this purpose, Spearman Correlation coefficients were calculated between each pair of features. A correlation coefficient between −0.75 and 0.75 was considered to indicate no significant correlation, while values between −0.75 and −1 or 0.75 and 1 were interpreted as indicating a significant correlation. The features identified as having a significant correlation based on these calculations are presented in Table 6.

Table 6. Correlations between selected attributes according to SHAP results obtained from trained Machine Learning models.

Based on the results of the correlation analysis, it was observed that there was no correlation among any of the features selected using SHAP values from the GB model. In contrast, there was a correlation between two pairs of features selected using SHAP values from the RF model, a correlation between six pairs of features selected using SHAP values from the RF LR model, and a correlation between seven pairs of features selected using SHAP values from the RF model.

Speers et al. reported that maternal embryonic leucine zipper kinase (MELK) ex-pression was significantly higher in breast cancer tissues compared to normal tissue and in TNBC compared to non-TNBC tissue, and the authors reported that MELK RNA and protein expression significantly correlated with radioresistance in BC cell lines (Xie et al., 2023). Another study reported that leucine deprivation inhibited cell proliferation and induced apoptosis of MDA-MB-231 and MCF-7 BC cells (Xiao et al., 2016). Singh et al. reported that leucine restriction was not sufficient to inhibit mammalian target of rapamycin (mTOR) signaling in most BC cell lines, but was associated with activation of the survival molecule Akt, making leucine deprivation an undesirable approach for BC therapy (Singh et al., 2011). In a biomarker discovery study, the amino acid Isoleucine (AUROC ≥ 0.85) was identified as a suitable candidate marker to diagnose and predict BC progression (Eniu et al., 2019).

The content of several intracellular branched-chain amino acids (BCAAs), including valine, leucine and isoleucine, has been reported to decrease after oridonin treatment. The results of this study suggest that oridonin has potent anti-tumor activity in vitro and in vivo and has potential as an adjuvant to BC treatment regimens. BCAAs are essential amino acids for the human body, and tumor tissues take up BCAAs from surrounding tissues or the bloodstream (Neinast et al., 2019). Therefore, high plasma levels of BCAAs have been as-sociated with cancer development (Mayers et al., 2014). Branched-chain amino acid transaminase 1 (BCAT1), which catalyzes the catabolism of essential BCAAs such as leucine, isoleucine and valine, plays an important role in amino acid metabolism. It can convert BCAAs into the corresponding branched-chain α-keto acids and then into α-ketoglutarate by transferring amino acids to produce glutamic acid, which can further promote the growth of BC cells (Zhang and Han, 2017). These branched-chain keto acids can be further oxidized to form acetyl-CoA and/or succinyl-CoA to feed the TCA cycle or contribute substrates to fatty acid synthesis (Mossmann et al., 2018). In a study that found that intracellular BCAT1 expression was significantly reduced after oridonin treatment, it was reported that this may have reduced intracellular nutrient and energy production. Since amino acids, including BCAAs, can also act as upstream regulators of mTOR activity (Jung et al., 2021), the authors also examined proteins involved in the mTOR signaling pathway. Western blot analysis revealed that the expression of mTOR, PI3K and AKT decreased significantly after oridonin treatment.

The biomarkers identified in the current study in BC management may not only facilitate treatment selection but also play an important role in clinical decision-making processes. Additionally, these molecular signatures may provide clinicians with a nuanced understanding of the heterogeneous nature of BC, allowing them to precisely tailor therapeutic interventions. Moreover, the biomarker-driven prediction model proposed in this study may be promising in determining prognosis by enabling the classification of patients into risk categories and optimizing therapeutic strategies. Consequently, inte-gration of biomarker profiling into routine clinical practice may be beneficial in the personalized management of BC and is important in targeted therapies and improved patient outcomes.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding authors.

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmolb.2024.1426964/full#supplementary-material

Arnold, M., Morgan, E., Rumgay, H., Mafra, A., Singh, D., Laversanne, M., et al. (2022). Current and future burden of breast cancer: global statistics for 2020 and 2040. Breast 66, 15–23. doi:10.1016/j.breast.2022.08.010

PubMed Abstract | CrossRef Full Text | Google Scholar

Bergstra, J., and Bengio, Y. (2012). Random search for hyper-parameter optimization. J. Mach. Learn. Res. 13, 281–305.

Google Scholar

Bhat, P. C., Prosper, H. B., Sekmen, S., and Stewart, C. (2018). Optimizing event selection with the random grid search. Comput. Phys. Commun. 228, 245–257. doi:10.1016/j.cpc.2018.02.018

CrossRef Full Text | Google Scholar

Bifarin, O. O. (2023). Interpretable machine learning with tree-based shapley additive explanations: application to metabolomics datasets for binary classification. Plos One 18, e0284315. doi:10.1371/journal.pone.0284315

PubMed Abstract | CrossRef Full Text | Google Scholar

Bonilla, J. M., Tabanera, M. T., and Mendoza, L. R. (2017). Breast cancer in the 21st century: from early detection to new therapies. Radiol. Engl. Ed. 59, 368–379. doi:10.1016/j.rx.2017.06.003

CrossRef Full Text | Google Scholar

Cao, W., Chen, H.-D., Yu, Y.-W., Li, N., and Chen, W.-Q. (2021). Changing profiles of cancer burden worldwide and in China: a secondary analysis of the global cancer statistics 2020. Chin. Med. J. 134, 783–791. doi:10.1097/CM9.0000000000001474

PubMed Abstract | CrossRef Full Text | Google Scholar

Dedeturk, B. K., and Akay, B. (2020). Spam filtering using a logistic regression model trained by an artificial bee colony algorithm. Appl. Soft Comput. 91, 106229. doi:10.1016/j.asoc.2020.106229

CrossRef Full Text | Google Scholar

DeSantis, C. E., Ma, J., Goding Sauer, A., Newman, L. A., and Jemal, A. (2017). Breast cancer statistics, 2017, racial disparity in mortality by state. CA Cancer J. Clin. 67, 439–448. doi:10.3322/caac.21412

PubMed Abstract | CrossRef Full Text | Google Scholar

Donisi, L., Cesarelli, G., Capodaglio, E., Panigazzi, M., D’Addio, G., Cesarelli, M., et al. (2022). A logistic regression model for biomechanical risk classification in lifting tasks. Diagnostics 12, 2624. doi:10.3390/diagnostics12112624

PubMed Abstract | CrossRef Full Text | Google Scholar

Eniu, D. T., Romanciuc, F., Moraru, C., Goidescu, I., Eniu, D., Staicu, A., et al. (2019). The decrease of some serum free amino acids can predict breast cancer diagnosis and progression. Scand. J. Clin. Lab. Invest. 79, 17–24. doi:10.1080/00365513.2018.1542541

PubMed Abstract | CrossRef Full Text | Google Scholar

Guo, Y., Zhang, Z., and Tang, F. (2021). Feature selection with kernelized multi-class support vector machine. Pattern Recognit. 117, 107988. doi:10.1016/j.patcog.2021.107988

CrossRef Full Text | Google Scholar

Hart, C. D., Vignoli, A., Tenori, L., Uy, G. L., Van To, T., Adebamowo, C., et al. (2017). Serum metabolomic profiles identify ER-positive early breast cancer patients at increased risk of disease recurrence in a multicenter population. Clin. Cancer Res. 23 (6), 1422–1431. doi:10.1158/1078-0432.CCR-16-1153

PubMed Abstract | CrossRef Full Text | Google Scholar

Hosseini, M., Powell, M., Collins, J., Callahan-Flintoft, C., Jones, W., Bowman, H., et al. (2020). I tried a bunch of things: the dangers of unexpected overfitting in classification of brain data. Neurosci. Biobehav. Rev. 119, 456–467. doi:10.1016/j.neubiorev.2020.09.036

PubMed Abstract | CrossRef Full Text | Google Scholar

Jasbi, P., Wang, D., Cheng, S. L., Fei, Q., Cui, J. Y., Liu, L., et al. (2019). Breast cancer detection using targeted plasma metabolomics. J. Chromatogr. B 1105, 26–37. doi:10.1016/j.jchromb.2018.11.029

PubMed Abstract | CrossRef Full Text |