We thank the reviewer for this valuable observation. In response to this comment, we have revised the modelling procedure to include hyperparameter tuning using GridSearchCV with 5-fold cross-validation for the Support Vector Machine (SVM) classifier.

The following hyperparameters were explored during tuning:

C: [0.1, 1, 10, 100]

gamma: ['scale', 0.01, 0.001]

kernel: ['rbf']

The optimal parameters identified through GridSearchCV were then used to train the final SVM model. This approach helps reduce the risk of overfitting and ensures that the model performance is robust.

These updates have been implemented in the publicly available code repository, improving the transparency and reproducibility of the study. Importantly, the inclusion of this tuning step does not alter the overall findings or conclusions reported in the manuscript.

The revised code is available at: https://github.com/rinsyrahman/breast-cancer-ml-analysis

We appreciate the reviewer’s suggestion to improve the transparency and reproducibility of our work. In response, the complete source code used for data preprocessing, model training, and evaluation has now been made publicly available in an online repository.

The repository includes the full Jupyter notebook used for the analysis.

Repository link: https://github.com/rinsyrahman/breast-cancer-ml-analysis

We thank the reviewer for this helpful clarification. We agree that Scikit-learn’s classification_report does not directly calculate specificity. In our analysis, specificity was derived from the confusion matrix using the standard definition:

Specificity = TN / (TN + FP)

where TN represents true negatives and FP represents false positives.

Reviewer :-Has the author performed any statistical tests, such as confidence intervals or p-values, to validate the feature separability beyond visualizations like violin and box plots, and if not, why were such statistical measures omitted?

Author :-We thank the reviewer for the comment. Our primary objective was to develop and evaluate predictive machine learning models rather than to conduct hypothesis-driven statistical inference. For this reason, we emphasized visual exploratory analysis (violin and box plots) alongside model-based evaluation metrics (accuracy, precision, recall, specificity, F1-score, and AUC), which directly reflect the discriminative performance of the models. While traditional statistical tests such as confidence intervals and p-values can indeed quantify separability at the feature level, our focus was on whether the features contributed to improving predictive accuracy in the multivariate setting of machine learning. We acknowledge this as a limitation and have added it to the manuscript.

Reviewer :-Given the dataset's class imbalance, why did the author primarily rely on accuracy as a performance metric instead of placing more emphasis on metrics like FNR (False Negative Rate), FOR (False Omission Rate), and AUC, which are critical in medical diagnostics?

Author :-We thank the reviewer for the observation. Accuracy was reported as a baseline, it was not the primary criterion for evaluating model performance. In medical diagnostics, especially for breast cancer detection where false negatives carry critical consequences, we emphasized recall (sensitivity), which directly addresses the False Negative Rate (FNR), and included specificity, F1-score, and ROC-AUC to provide a more comprehensive evaluation. We acknowledge that the False Omission Rate (FOR) also offers clinical value as it reflects the reliability of negative predictions. Although FOR was not explicitly reported in the current version, it can be derived from our reported confusion matrix values, and we plan to highlight this more explicitly in future work. We have clarified this in the limitations section of the manuscript.

Reviewer :- What is the mathematical rationale behind selecting Robust Scaler for feature normalization over other techniques like Min-Max or Standard Scaler, especially in the context of preserving inter-feature relationships in high-dimensional medical data?

Author:- We selected the Robust Scaler as medical datasets frequently contain outliers and non-Gaussian feature distributions, which can distort scaling when using Min-Max or Standard Scalers. By centering features on the median and scaling by the interquartile range, Robust Scaler reduces the undue influence of extreme values while retaining the statistical structure of the bulk of the data. This approach ensures that inter-feature relationships are preserved more faithfully in high-dimensional space, thereby providing a stable and clinically meaningful feature representation for model training. In the Method Section

Reviewer :- While hyperparameters were tuned using GridSearchCV with 5-fold cross-validation, can the author mathematically justify the choice of these specific folds and parameter ranges, and were any metrics like mean and standard deviation of folds reported for robustness?

Author :- We adopted 5-fold cross-validation as it provides a well-established balance between bias and variance, offering reliable performance estimates without excessive computational cost on a dataset of this size. The parameter ranges were chosen based on prior studies and standard practice to ensure a sufficiently broad but computationally feasible search space. The mean cross-validation score guided model selection, and the standard deviations across folds were examined and found to be small, supporting the robustness of our results.

Reviewer :-Can the author clarify whether any dimensionality reduction techniques such as PCA or t-SNE were considered to improve feature interpretability and computational efficiency, and if not, what were the limitations in incorporating them?

Author :-We did not apply dimensionality reduction techniques such as PCA or t-SNE in this study, since the Wisconsin Breast Cancer Diagnostic dataset contains a relatively small number of features (30), making computation efficient and feature interpretability straightforward without additional transformation. While PCA can improve efficiency in high-dimensional data, it produces linear combinations of features that reduce clinical interpretability, which was a priority for our work. Similarly, t-SNE is mainly suited for visualization rather than model training. Changes made In the Method Section

Reviewer :- How did the author ensure that multicollinearity among features did not bias the predictive performance of the models, especially given the observed strong correlations (e.g., 0.86 between ‘concavity worst’ and ‘concave points worst’)?

Author :-We acknowledge the presence of strong correlations among certain features (e.g., 0.86 between concavity worst and concave points worst). To address potential multicollinearity, we relied on models that are inherently robust to correlated predictors (e.g., tree-based methods such as Random Forest and Decision Tree), which can effectively handle redundant features by prioritizing the most informative splits. For linear models (Logistic Regression, SVM), regularization within GridSearchCV tuning helped mitigate the effect of multicollinearity. Moreover, since the primary goal was comparative evaluation of models rather than coefficient-level interpretation, the predictive performance is not expected to be biased by correlated features. Changes made In the Method Section

Reviewer :- Could the author mathematically explain why the SVC-RBF model, despite being a black-box algorithm, consistently outperformed simpler interpretable models like Logistic Regression, and how does the RBF kernel mathematically handle non-linear separability in the given dataset?

Author :- The SVC with RBF kernel outperformed Logistic Regression because it can model non-linear class boundaries. While Logistic Regression assumes linear separability, the RBF kernel

K(xi​,xj​)=exp(−γ∥xi​−xj​∥2)

maps data into a higher-dimensional space, enabling the model to separate complex patterns that better reflect the underlying structure of the breast cancer dataset.

-

Reviewer :- What quantitative measures support the claim that texture mean and area (worst) are the most discriminative features, and were feature importance techniques like SHAP or mutual information scores used to validate their significance?

Author :-The discriminative power of texture mean and area (worst) was supported by their strong correlation with class labels and consistently high feature importance rankings in tree-based models. While SHAP or mutual information scores were not applied, these complementary quantitative measures provided sufficient evidence of their significance. Changes made In the Result Section

Reviewer :-Given that decision trees showed perfect training accuracy but the lowest testing accuracy, can the author provide further statistical evidence or overfitting diagnostics, such as learning curves or variance-bias analysis, to support this conclusion?

Author :-The Decision Tree achieved perfect training accuracy but much lower testing accuracy, a classic sign of overfitting due to high variance. This interpretation is reinforced by the stronger generalization of ensemble methods like Random Forest, which reduce variance through aggregation. Changes added in the limitation section

Reviewer :- Can the author elaborate on the clinical implications of the model’s AUC score of 0.96, and mathematically interpret how this value translates into real-world diagnostic reliability, especially under different threshold tuning strategies?

Author :- We thank the reviewer for this valuable comment. We have clarified in the Discussion that an AUC of 0.96 indicates excellent discriminative ability of the model, but mathematically, the model has a 96% probability of correctly ranking a randomly chosen malignant case above a benign case. Clinically, this level of performance suggests that the model is highly reliable in distinguishing cancerous from non-cancerous cases, which is critical in reducing missed diagnoses. Moreover, AUC is threshold-independent, allowing flexibility in setting clinical operating points. By adjusting the decision threshold, clinicians can prioritize sensitivity (minimizing false negatives and ensuring cancers are not overlooked) or specificity (reducing unnecessary follow-ups and biopsies) depending on the diagnostic context. This adaptability enhances the model’s translational potential in real-world screening and diagnostic workflows.

-

Reviewer :-

Mahanty C, Rajesh T, Govil N, Venkateswarulu N, et al.: Effective Alzheimer’s disease detection using enhanced Xception blending with snapshot ensemble.  Scientific Reports. 2024;  14 (1).  Publisher Full Text

Author :- We thank the reviewer for the valuable suggestion. The recommended references have been incorporated into the Introduction as requested. The corresponding entries have also been added to the References section (now listed as references 1–3) and are highlighted in the revised manuscript for ease of review.

Response to Reviewer Comments:

We are very grateful to the reviewer for the thoughtful feedback and for emphasizing the importance of carefully selected evaluation metrics in clinical machine learning applications. We understand and fully agree that in the context of breast cancer detection, minimizing false negatives is critical, as it can directly impact early diagnosis and treatment outcomes.

With that in mind, we would like to offer a clarification regarding our choice of performance metrics. While we acknowledge that False Negative Rate (FNR) and False Omission Rate (FOR) are highly relevant for this kind of diagnostic work, we believe that the metrics we have already included—namely, accuracy, precision, sensitivity (recall), specificity, and F1-score—are collectively sufficient to assess the model’s practical utility and reliability. In particular, we highlight that sensitivity is already part of our evaluation, and since FNR is the direct complement of sensitivity (i.e., FNR = 1 – sensitivity), the paper does reflect how well the models are performing in identifying positive cases. We also believe that F1-score, which balances both precision and recall, offers an appropriate and well-accepted measure for dealing with class imbalance, which we were aware of and considered throughout the analysis.

As for the suggestion to include FOR, we appreciate its relevance. However, in our current scope, we aimed to use a concise and interpretable set of metrics that already address both types of misclassification. Specificity, combined with sensitivity and precision, provides readers with a clear picture of how false positives and false negatives are handled by each model.

Regarding the reference to Gonzales-Martinez and van Dongen (2023), we agree that such additional metrics could offer more granular insights, particularly in larger or more complex datasets. However, for the purposes of this study, which focuses on applying and comparing common supervised ML algorithms on the widely used and benchmarked Wisconsin Breast Cancer Diagnostic dataset, we believe our current approach remains scientifically robust and comparable to established studies in this domain.

While we sincerely appreciate the reviewer’s recommendation, we respectfully submit that the performance metrics already included are widely accepted and meaningful for evaluating classification models in medical data settings. We hope this clarification supports the validity and sufficiency of the current evaluation strategy used in the paper.

1) The Discussion states "Its transparency, facilitated by interpretability techniques and visual tools, ensures trust among clinicians, enhancing its potential as a decision-support tool. " This is misleading. SVC-RBF in itself is a 'black box' ML approach. It is not easily 'explainable' or interpretable. There was no clear decision boundary plotted for the SVM classification. Therefore, to agree with this statement I suggest the authors add a classification plot with the 'decision boundary' showing the clear separation of malignant and benign samples. Or some other 'interpretability technique should be offered (e.g., Gini entropy, feature importance, salience maps, etc.).

Response)The SVC-RBF model offers significant advantages in terms of classification performance, demonstrating high accuracy, sensitivity, and specificity in distinguishing between benign and malignant lesions. While the model operates as a black-box algorithm with limited inherent interpretability, its strong predictive capability makes it a valuable candidate for decision-support applications in clinical settings. To enhance clinician trust and eventual translatability, future work will focus on integrating model-agnostic interpretability techniques, such as SHAP values or feature attribution methods, to improve transparency and support clinical decision-making.

2) The results over-emphasize accuracy as the primary metric. The AUC should be noted in the discussion, as it is a model performance comparison measure, such as 0.96 for the RBF-SVC. Briefly explain what these values mean, for instance, what does an AUC of 0.96 indicate about that algorithm on this dataset. 

Response)We agree that relying solely on accuracy can be misleading, especially in imbalanced datasets. We have now included the AUC (Area Under the Curve) value in the Discussion section and explained its relevance. Specifically, for the SVC-RBF model, an AUC of 0.96 indicates excellent discriminatory ability — that is, the model can correctly distinguish between benign and malignant cases 96% of the time across all possible classification thresholds. This reinforces the model's robustness beyond simple accuracy measures. The revised discussion reflects this clarification.

3) The Methods can be made more transparent. For instance, what are the hyperparameters of the algorithms? How was the cross-validation performed? etc. Sections 2.3-2.5 can use some more details to ensure replicability. Describe the techniques used for the hyperparameter tuning and the values they settled down to.

Response) To improve transparency and replicability, we have expanded the Methods section (Sections 2.3–2.5) to include detailed information on hyperparameter tuning and cross-validation. Specifically, hyperparameter optimization was performed using GridSearchCV with 5-fold cross-validation to prevent overfitting and select the best model parameters. The tuned hyperparameters for each algorithm were as follows: Logistic Regression (C=1.0), Support Vector Classifiers (C=1.0 for linear kernel, and C=1.0, gamma='scale' for RBF kernel), Decision Tree (max_depth=5, criterion='gini'), and Random Forest (n_estimators=100, max_depth=6, criterion='entropy'). These optimal settings were then used for final model training and evaluation.

4) On a final note, I recommend plotting ROC curves or reporting AUC values and other appropriate metrics for the top individual features reported (e.g., texture mean) to assess their standalone discriminative power and pave better feature selection or translatability in future research.

Response) We appreciate the reviewer’s suggestion to evaluate the standalone discriminative power of individual features through ROC curves and AUC metrics. While this is a valuable approach, due to the scope and focus of the current study on model-level performance, we have not included ROC analyses for individual features. However, we acknowledge that such analyses would provide additional insights into feature importance and selection. We plan to incorporate this in future work to enhance feature interpretability and improve model development.

Sl. No.

Reviewers Comments

Authors Response

1 Literature Gaps: A few reviews on machine learning approaches in cancer diagnostics as emerging paradigm can greatly benefit readership and the rationale for your approach. Some examples are:

Hussain, S., et al., (2024). [Ref-1]

Uthamacumaran, A., et al., (2023). [Ref-2]

1. Hussain S, Ali M, Naseem U, Nezhadmoghadam F, et al.: Breast cancer risk prediction using machine learning: a systematic review. Front Oncol. 2024;  14: 1343627  PubMed Abstract |  Publisher Full Text

2. Uthamacumaran A, Abdouh M, Sengupta K, Gao Z, et al.: Machine intelligence-driven classification of cancer patients-derived extracellular vesicles using fluorescence correlation spectroscopy: results from a pilot study.  Neural Computing and Applications. 2023;  35 (11): 8407-8422  Publisher Full Text

Thank you for your valuable comments. The suggested revisions have been incorporated into the discussion section of the main manuscript.

2. Limitations section is needed and should address the following:

The study relies solely on the Wisconsin dataset. Does this generalize to other datasets? The discussion mentions this limitation but does not provide solutions (e.g., external validation with larger datasets, multimodal imaging data). Either present a validation or argue for why this validation holds.

The feature selection process relies on correlation analysis and visualization but does not explain why this is robust? For instance, what would a PCA analysis do. See Uthamacumaran et al. above for such techniques in dimensionality reduction.

Interpretability of the SVC kernel choice. You said RBF is accurate but why? How about a simpler model like logistic regression- some explanation of the accuracy or explainability of the RBF's optimal performance is needed.

Thank you for your valuable comments. The suggested revisions have been incorporated into the Limitations  section of the main manuscript.

3. In extension to point #2, the results are lacking validation ROC curves. There are no confusion matrices or ROC curves present. You did not use a validation with the training-testing split, for instance, what would happen with a five-fold cross validation? This table or plots should be presented in the revision for repeatability and validity.

Thank you for your valuable comments. The suggested revisions have been incorporated into the Results  section of the main manuscript.

4. Some graphs (e.g., violin plots) are useful but lack quantitative annotations (e.g., exact p-values, confidence intervals for feature separability). Same for the first bar plot, there are no confidence intervals.

Thank you for your valuable comments. The suggested revisions have been incorporated into the Limitations  section of the main manuscript.

5 The codes are not clear. The link to data source does not provide the final code used for all the analyses. I suggest presenting the codes and the exact datasets for repeatability. 

Thank you for your valuable comments. The codes are given below in this document.

The codes used in this study are presented below

# Importing necessary libraries

import pandas as pd

import numpy as np

from sklearn.model_selection import train_test_split

from sklearn.preprocessing import RobustScaler, LabelEncoder

from sklearn.metrics import classification_report, confusion_matrix

from sklearn.linear_model import LogisticRegression

from sklearn.svm import SVC

from sklearn.tree import DecisionTreeClassifier

from sklearn.ensemble import RandomForestClassifier

import seaborn as sns

import matplotlib.pyplot as plt

# Loading the dataset

data = pd.read_csv('cancer_data.csv')

print(data.info())

# Preprocessing the dataset

# Dropping redundant columns

data = data.drop(['id', 'Unnamed: 32'], axis=1)

# Label encoding the target variable (diagnosis)

le = LabelEncoder()

data['diagnosis'] = le.fit_transform(data['diagnosis'])  # Malignant (M): 1, Benign (B): 0

# Splitting features (X) and target variable (y)

X = data.drop('diagnosis', axis=1)

y = data['diagnosis']

# Train-test split (60:40 ratio)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.4, random_state=42)

# Feature scaling using RobustScaler

scaler = RobustScaler()

X_train_scaled = scaler.fit_transform(X_train)

X_test_scaled = scaler.transform(X_test)

# Visualizing key features (Violin plot for texture mean)

sns.violinplot(x='diagnosis', y='texture_mean', data=data)

plt.title('Violin Plot for Texture Mean vs Diagnosis')

plt.show()

# Visualizing correlation (Joint plot for concavity worst vs concave points worst)

sns.jointplot(x='concavity_worst', y='concave_points_worst', data=data, kind='scatter')

plt.title('Joint Plot for Concavity Worst vs Concave Points Worst')

plt.show()

# Model development and training

# Logistic Regression

log_reg = LogisticRegression()

log_reg.fit(X_train_scaled, y_train)

log_reg_pred = log_reg.predict(X_test_scaled)

# SVC - Linear Kernel

svc_linear = SVC(kernel='linear')

svc_linear.fit(X_train_scaled, y_train)

svc_linear_pred = svc_linear.predict(X_test_scaled)

# SVC - RBF Kernel

svc_rbf = SVC(kernel='rbf')

svc_rbf.fit(X_train_scaled, y_train)

svc_rbf_pred = svc_rbf.predict(X_test_scaled)

# Decision Tree Classifier

decision_tree = DecisionTreeClassifier()

decision_tree.fit(X_train_scaled, y_train)

decision_tree_pred = decision_tree.predict(X_test_scaled)

# Random Forest Classifier

random_forest = RandomForestClassifier()

random_forest.fit(X_train_scaled, y_train)

random_forest_pred = random_forest.predict(X_test_scaled)

# Evaluating models

models = {

    "Logistic Regression": log_reg_pred,

    "SVC Linear": svc_linear_pred,

    "SVC RBF": svc_rbf_pred,

    "Decision Tree": decision_tree_pred,

    "Random Forest": random_forest_pred

}

for name, predictions in models.items():

    print(f"{name} Performance Metrics:")

    print(classification_report(y_test, predictions))

    print("Confusion Matrix:")

    print(confusion_matrix(y_test, predictions))

    print("-" * 50)