TURKISH JOURNAL OF ONCOLOGY

Summary

OBJECTIVE
The necessity to diagnose breast cancer early and correctly is the need to minimize the diagnostic uncertainty and unwarranted clinical procedures. This paper assesses the reliability of a hybrid deep-ensemble decision-support model in terms of diagnostic reliability, stability of outcome, and translational feasibility of the model via structured clinical data to detect early breast cancer.

METHODS
The Wisconsin Diagnostic Breast Cancer dataset which consisted of 569 cases of benign and malignant tumors was analyzed retrospectively. The framework proposed combines the deep learning of latent representations with stacked classification, ensemble-based feature selection, and stacked classification. Performance evaluation was performed based on sensitivity, specificity, accuracy, F1-score, and area under the curve (AUC) performed using stratified 10-fold cross-validation. The statistical stability across folds and the comparison with baseline models were determined with the help of non-parametric tests (p<0.05).

RESULTS
The model had good diagnostic performance with an accuracy of between 91.2-100 (Mean 96), Sensitivity of 76.2-100, good specificity value, and AUC 0.973-1.000. Variability in performance between folds was low, and statistically significant enhancement as compared to baseline classifiers were present.

CONCLUSION
The hybrid deep-ensemble model is highly diagnostic, has robust discriminative ability, and ultimately remains stable, which demonstrates the methodological robustness and diagnostic reliability of the proposed framework as a proof-of-concept decision-support model for early breast cancer detection, with potential translational relevance subject to further external clinical validation.

Summary

Introduction

Breast cancer is the second leading cancer killer of women in the world and as such, there is the need to ensure that cancer is detected at early age to have better outcomes. Digital Mammography (DM) has the potential to save 40% of death due to breast cancer. Nevertheless, among patients who have high density of the breast, small early tumors are more problematic that results in the rise in false positives and interval malignancies. Currently 43 percent of women between the ages 40-85 years are having thick breasts and this implies that they require other forms of screening besides DM Kinkar et al.[1] which include X-ray mammography, ultrasound, Computed Tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI) and breast temperature. The gold standard of diagnosis is the pathological examination of excised tissue that is complemented by staining (heatsylin and eosin (H&E)). It can be diagnosed by a study with the assistance of histological images or genomics as breast cancer is the cause of death of most women because it is sensitive to various clinical, lifestyle, social, and economic factors. Such techniques as MRI, X-ray, and mammograms can also be used with the help of machine learning, particularly, Convolutional Neural Networks (CNN), in the early diagnosis.[2] The examination of the Wisconsin Breast Cancer Daterset revealed that CNN is highly efficient when it comes to detecting the presence of breast cancer, and the most significant tool of evaluation is accuracy measures. The application of deep learning to real life is expanding quickly.[3] Breast cancer detection and diagnosis with segmentation models with the primary focus on the Unet3+ one.[4] It compares Unet3+ with other models like Fully Convolutional Networks (FCN), Unet, SegNet, Deeplab V3+ and pspNet using images of 309 patients (151 benign and 158 malignant). Unet3+ had the highest accuracy mean score; this is 82.53 and global accuracy; this is 90.99. The authors introduce the Efficient Adaboost Algorithm (DLA-EABA) of breast cancer detection using Deep Learning Zheng et al.,[5] where neural networks (in particular, CNNs) would be used in the process of tumour classification. It integrates machine learning and feature selection and extraction, accuracy of 97.2, sensitivity of 98.3 and specificity of 96.5 which are higher than the past systems. The dataset was considered as 3002 computerized mammograms of 1501 cancer patients with diverse density to train a deep learning model that was used to predict breast cancer in computerized mammograms with diverse density with low computing requirements.[6] Classification frameworks, including random forest, decision tree, k-nearest neighbors, logistic regression, support vector classifier and linear support vector classifier, were adopted, which were quite efficient, precise and minimum computing demands. The risk of breast cancer is predicted to have an accuracy of 91 percent, according to one of the studies, deep learning model based on transfer learning and InceptionResNetV2, which is developed to improve the risk assessment factors, opens a prospect of turning nearly all medical imaging processes into an automated process.[7] The recent studies indicate that the sphere of breast cancer diagnosis by means of the methods of machine learning(ML) and deep learning(DL) is vastly underdeveloped yet there exist several disadvantages between the existing methods.[8-12] Very high-quality handcrafted ultrasound features can be found in optimized ML classifiers, such as LightGBM, but these models do not have much deep representational capacity and require large data sets, large-scale computation, and lack much interpretability.[13] Mammographic preprocessing and texture-based SVM-ANFIS Kumar et al.[14] classification with a high degree of accuracy but highly dependent on image processing and analysis that involve various handcrafted processes, which suppresses scalability.[15] Strong performance of ensemble-based models is due to the use of only a classic ML, with no deep feature extractors or hybrid. It is important to note that this study is positioned as a methodological proofof- concept, aiming to evaluate diagnostic reliability, stability across validation folds, and false-positive control using structured clinical diagnostic features. While the proposed framework is motivated by clinical decisionsupport needs, direct deployment in population-level screening workflows would require additional validation on prospective, multi-center, and imaging-based datasets. Accordingly, clinical applicability claims in this work are framed in terms of translational potential rather than immediate clinical adoption (Table 1).

Table 1 Comparative summary of existing breast cancer detection studies and the proposed HDEL-BC framework

Problem Statement, Research Gap and Motivation
The early identification of breast cancer is a vital factor to achieve better patient results; nevertheless, the current computational diagnostic strategies tend to focus on the independent increase of accuracy at the cost of clinical reliability, stability, and control of false positives. Most papers are either based on handcrafted functionality, deep learning models that are computationally intensive or, are based on single-stage classifiers, thereby restricting interpretability, generalization and real-world implementation in routine clinical environments. Furthermore, there has been less focus on outcome stability and statistical strength across validation folds diminishing translational confidence. These gaps have inspired this study to create computationally efficient, clinically reliable, and interpretable diagnostic decision-support structure, which is driven by deep latent representations and by optimizing robustness and reducing false-positives, and can be readily deployed in resource-constrained clinical settings.

Novelty and Contributions of the Study
This paper introduces a single-user-friendly hybrid deep-ensemble decision-support system (HDEL-BC) to identify breast cancer early in its progression with the focus on diagnostic accuracy, stability, and clinical relevance and not individual accuracy improvement. In the proposed method, the latent features obtained by the autoencoders are combined with clinically interpretable standardized diagnostic variables to create a hybrid feature space that balances predictive accuracy with interpretability. The framework maintains practical interpretability at the decision-support level by integrating deep latent representations with clinically interpretable diagnostic features. To enhance the stability of features and minimize redundancy, the three-ensemble feature selection strategy (LASSO, Random Forest, and XGBoost) is used, and the heterogeneous stacked ensemble classifier is applied to allow consensus decision making. The framework has been shown to have a very high specificity in stratified cross-validation, low false-positive rates, and outstanding discriminative performance, and its statistical analysis has shown significant distance superiority to distance-based classifiers and similar performance to other strong baseline models. In general, the work provides a computational efficient and clinically reliable diagnostic framework that can be employed to screen breast cancer at the early stage and applied into the clinical environment with limited resources available.

Introduction

Methods

Study Design
The current study was a case of building a retrospective diagnostic outcome study in order to investigate the reliability and translational usefulness of a cross-breeding profound-ensemble decision support framework concerning early detection of breast cancer. The study was performed as a secondary use of an open, anonymized dataset and did not imply any direct contact with the patient, clinical intervention and specific health data. Thus, there was no need of formal ethical approval and informed consent.

Study Population and Dataset
The dataset used in the study was the Wisconsin Diagnostic Breast Cancer (WDBC) dataset that included 569 cases of diagnosis classified as either benign or malignant according to the histopathological results. Both cases are defined by 30 quantitative findings of diagnosis on the basis of digitized images of the fine-needle aspirate (FNA) of the breast masses.[16]

These characteristics reflect clinically significant morphological attributes, such as size of tumor, texture, perimeter, area, concavity, symmetry, and fractal attributes. These qualities are habitually put into consideration when diagnostic assessment is carried out and offer a proper baseline when automated diagnostic decision-support systems are assessed.

Data Preprocessing and Bias Mitigation
In order to assure numerical stability and comparability of all features, all diagnostic variables were z-score standardized, which is defined as:

where ×_i denotes the original feature value, and µ and ó represent the mean and standard deviation computed from the training data, respectively. The Synthetic Minority Over-Sampling Technique (SMOTEENN) was used to deal with the question of class imbalance between benign and malignant cases. Notably, crossvalidation was done via standardization and resampling of the respective train data only in the training data of every fold of the cross-validation process without data leakage or optimistic bias. Tests folds were completely independent during the model development. Although SMOTEENN introduces synthetic minority samples during training, its use in this study is restricted exclusively to the training folds within each cross-validation iteration, thereby preventing information leakage into the test data. The objective of resampling is not to model biological variability directly, but to reduce classifier bias arising from class imbalance and to stabilize decision boundaries during learning. Since all performance metrics are computed on original, non-synthetic samples, the reported diagnostic outcomes reflect real-case behavior rather than synthetic artifacts.

Hybrid Deep-Ensemble Diagnostic Framework
The suggested diagnostic framework in Figure 1 combines deep latent representation learning with ensemblebased feature selection and stacked classification so that to improve diagnostic robustness and generalization.

Fig. 1. Conceptual framework of the proposed HDEL-BC decision-support system, illustrating how structured clinical diagnostic features are transformed into stable malignancy risk estimates through hybrid learning and ensemble decision aggregation to support early breast cancer assessment.

Deep Latent Representation Learning: A small autoencoder neural network was used to train nonlinear latent features on standardized diagnostic features. It was based on an encoder-decoder architecture that had an 8-dimensional bottleneck layer, which allowed the feature of high dimension in the input to be compressed into a low-dimensional latent space. The autoencoder was optimized using mean squared reconstruction error so that latent features can be extracted that reproduce a greater number of diagnostic relationships.

Hybrid Feature Construction: The extracted latent features were also combined with the original standardized diagnostic features to form a hybrid feature representation to provide a trade-off between the predictive performance and the clinical interpretability. Such a strategy does not only retain explicit morphological data, but also adds nonlinear diagnostic patterns to the feature space. While the autoencoder-derived latent features are not individually interpretable, their integration with original diagnostic variables ensures that clinically meaningful features continue to contribute to the final decision-making process.

Ensemble-Based Feature Selection: Since the hybrid feature set was increased, tri-ensemble feature selection strategy was utilized in determining stable and diagnostically telling features. The importance of feature was estimated independently as:

1. Regularized L1-regularized logistic regression,

2. Random Forest, importance, based on impurity, And

3. Gradient boosting (XGBoost) gain measures.

The results of these methods were normalized and the results were combined to produce a consensus ranking and the features that were ranked highly were then chosen to be used in further classification. This is an ensemble method that minimizes the bias of each model and enhances stability of features.

Stacked Classification Strategy: The prediction of final diagnosis was done with the help of a stacked ensemble classifier, which consisted of heterogeneous base learners such as gradient boosting, random forest, support vectors machine, and multilayer perceptron models. A logistic regression meta-learner was used to combine the probabilistic outputs of the base classifiers so as to estimate the final probability of malignancy:

where represents the probability predicted by the

base classifier, denotes the learned weight, and ó(.) is the sigmoid activation function.

Model Validation Strategy
At one time, stratified 10-fold cross-validation was used to assess model performance, making sure that the distribution of classes remains the same in every fold. Training and independent testing one-fold was applied in every iteration. The whole process of pre-processing, feature learning, feature selection, and model training was performed only in the training folds giving an impartial approximation of the diagnostic performance.

Outcome Measures
The diagnostic performance was evaluated by clinically relevant outcome measures, which were sensitivity, specificity, accuracy, F1-score, and the area under the receiver operating characteristic curve (AUC). These measures were stipulated as:

Through means of AUC, a threshold-independent measure of discriminative performance was conducted.

Statistical Analysis
The results of the performance measures were summarized by cross-validation in terms of the means and variability estimates. 95% confidence intervals were calculated where possible. Non-parametric statistical testing was used to compare the proposed framework and baseline classifiers and statistical significance was set at two-sided p-value value smaller than 0.05. The implementation of all analyses was done based on standard Python-based scientific computing libraries. The materials and methods were formulated to allow an objective assessment, statistical soundness, and clinical significance, therefore, choosing the applicability of the proposed decision-support framework to translational applications.

Methods

Results

Study Population and Class Distribution
An experimental assessment was undertaken on 569 cases of breast cancer of which 357 (62.7%) were benign and 212 (37.3%) were malignant cases. The class distribution of a clinically realistic diagnostic scenario where the benign findings are higher than the malignant cases, indicating the necessity of the false-positive control of automated diagnostic systems.

Cross-Validated Diagnostic Performance
The proposed hybrid deep -ensemble learning (HDELBC) model was tested based on 10-fold cross-validation with stratification to provide robust and unbiased estimations of the model performances. Table 2 gives the fold-wise diagnostic measures.

Within the validation folds, accuracy varied between 91.2% and 100% with most of the folds having a high value of above 96, which is an indication of consistent generalization. The sensitivity had a range of 76.2% to 100% which indicated conservative behavior in some folds whereas the specificity was always high and in some cases it was 100%. The curve below ROC (AUC) was between 0.973 and 1.000 which validates high discriminative ability.

Table 2 Fold-wise cross-validated diagnostic performance

Performance Stability Across Validation Folds
Boxplots of AUC, sensitivity, specificity, and accuracy are used to show the consistency of diagnostic performance between the validation folds. Figure 2 shows the stability of the performance of the proposed framework when cross-validated over 10 folds. The AUC is also not changing widely, it is staying relatively constant at 0.97-1.00, which represents a good discriminative ability. Tight distributions are also shown by specificity and accuracy, most values are above 0.95, indicating a good control of false-positives and the general correctness. Sensitivity shows a little higher variation between 0.76 and 1.00 indicating conservative capture behavior at some of the folds.

The fact that the dispersion is still low in all the measures, especially in the case of AUC and specificity, shows that it is highly robust and does not seem to rely on a particular data partition. This is due to the fact that the relatively greater spread in sensitivity indicates conservative detection behavior of certain folds, which is more clinically desirable than over-reporting false-positives.

Fig. 2. Cross-validated performance stability across 10 folds, highlighting consistent specificity and low variability in diagnostic outcomes, which are critical for reducing false-positive findings in clinical breast cancer screening.

Receiver Operating Characteristic Analysis
To further investigate the discriminative behavior of the proposed framework, receiver operating characteristic (ROC) curves were also constructed in each of the validation folds. The curves of receiver operating characteristics (ROC) that were obtained during 10-fold cross-validation are shown in Figure 3. The curves are continuously directed towards the upper-left corner, which shows that there is a high level of discriminative performance between folds. The values of the AUC corresponding to 0.97 to 1.00 show that the true-positive rates are high at very low false-positive rates. This agreement between folds confirms the strength and trustworthiness of the coined framework in making out the differences between malignant and benign cases.

The ROC curves are always near to the upper-left part, which proves the high true-positive values at low false-positive ones and proves the high level of discrimination of the model.

Fig. 3. Receiver operating characteristic (ROC) curves across 10-fold stratified cross-validation, demonstrating reliable discrimination between benign and malignant cases while maintaining low falsepositive rates relevant to clinical decision support.

Aggregated Error Analysis and Clinical Safety
The confusion outcomes were summed up in Table 3 to evaluate the overall diagnostic safety. This model was able to yield 350 true negatives and 200 true positives and only 7 false positives and 12 false negatives.

Table 3 Aggregated confusion matrix outcomes

The low percentage of false-positive (1.96) in Table 2 is especially useful in breast cancer screening as it reduces false biopsy, extra examination, and patient anxiety. The false-negative rate that was observed is a tradeoff between specificity and sensitivity of a conservative diagnostic system.

Statistical Comparison with Baseline Models
The proposed framework was statistically compared to conventional baseline classifiers in terms of diagnostic performance using the Wilcoxon signed-rank test of the AUC values and fold-wise results and a table was provided in Table 4. Although the Wilcoxon signed-rank analysis shows that several strong baseline classifiers achieve statistically comparable AUC values, this does not diminish the practical contribution of the proposed framework. The primary objective of HDEL-BC is not to maximize marginal performance gains, but to achieve stable diagnostic behavior, low false-positive rates, and consistent specificity across validation folds. Such characteristics are particularly relevant in clinical decision-support settings, where reliability and safety often outweigh small differences in aggregate performance metrics.

Table 4 Wilcoxon signed-rank test results (HDEL vs Baseline Models)

The analysis indicates that there is statistically significant enhancement over the distance-based K-Nearest Neighbors (KNN) classifier (p<0.01) but the performance was statistically the same as the other robust baseline classification models, such as, logistic regression, random forest, support vector machine, and XGBoost.

These results demonstrate strong discriminative performance under cross-validated benchmark conditions, while acknowledging that performance on curated datasets may not directly translate to real-world clinical settings.

Discussion and Clinical Interpretation
The findings show that the hybrid deep-ensemble model proposed has good and consistent diagnostic accuracy in the early detection of breast cancer when subjected to stringent cross-validation. Instead of concentrating only on marginal improvements in accuracy, the framework concentrates on the outcome stability, false-positive suppression, and clinical safety that are important factors in the application of diagnostic decision support in real-world settings.

From a clinical perspective, the use of synthetic resampling techniques is primarily intended to support balanced learning rather than to simulate biological heterogeneity. The consistently high specificity and low false-positive rates observed across validation folds suggest that the inclusion of synthetic minority samples did not result in overly aggressive sensitivity or unsafe diagnostic behavior. Instead, the framework exhibits conservative decision characteristics, which are generally preferable in breast cancer screening contexts where false-positive reduction is a critical concern.

The fact that the latent deep representations are combined with the clinically interpretable features is a feature that allows the identification of the nonlinear pattern of diagnosis without the loss of transparency. The ensemble feature selection and stacked classification also lead to less variability of performance in the different validation folds. Notably, the framework can attain these results without using raw imaging data or architectures that require extensive computation, which makes it plausible to implement it in resource-limited clinical settings. While the proposed architecture introduces additional components compared to single-model baselines, this complexity is justified by improved outcome stability and conservative diagnostic behavior rather than by isolated accuracy improvements. The consistently high accuracy and AUC values observed across validation folds should be interpreted with caution. The WDBC dataset is a well-curated and widely used benchmark, and strong performance may partly reflect dataset saturation rather than generalizable real-world diagnostic behavior. Although stratified cross-validation and strict train-test separation were employed to mitigate overfitting, external validation on independent and heterogeneous clinical datasets is necessary to confirm generalizability and clinical robustness.

Despite the encouraging results, this study has several important limitations. First, the evaluation is restricted to the WDBC dataset, which is a curated benchmark dataset derived from fine-needle aspirate-based diagnostic features rather than population-level screening or raw imaging data. Consequently, the findings should be interpreted as proof-of-concept evidence demonstrating diagnostic stability and robustness under controlled validation settings. External validation using multi-institutional datasets, prospective cohorts, and imaging-based screening data is required before the framework can be considered for real-world clinical deployment. Nevertheless, the present results establish a reliable methodological foundation for such future translational studies.

The hybrid deep-ensemble framework that is proposed in this paper shows a good, consistent, and clinically viable performance in detecting breast cancer in its early stages with a high level of discrimination, low false-positive probability, and statistically stable comparative performance, which adds evidence to the translational viability in the use of the framework as a decision-support system in breast cancer diagnosis. Although the proposed framework emphasizes interpretability at the system and feature-integration level, no explicit feature-attribution or explainability analysis of individual latent dimensions was performed, which represents an important direction for future work.

Results

Conclusion

Early and reliable identification of breast malignancy remains central to improving patient outcomes and optimizing downstream diagnostic pathways. In this study, a hybrid deep-ensemble decision-support framework (HDEL-BC) was developed and evaluated with a specific focus on diagnostic safety, stability, and clinical usability rather than isolated performance optimization. By integrating autoencoder-derived latent representations with standardized, clinically interpretable diagnostic features, the proposed approach provides a balanced representation of both hidden morphological patterns and established radiological descriptors relevant to breast cancer assessment.

The framework demonstrated consistently high specificity, low false-positive rates, and strong discriminative ability across stratified cross-validation, supporting its potential role as a screening-support tool aimed at minimizing unnecessary follow-up procedures and patient anxiety. Statistical analysis further indicated robust performance relative to commonly used baseline models, reinforcing the reliability of the framework under repeated validation rather than dependence on single-split performance.

From an oncology perspective, the proposed model is computationally efficient and interpretable, making it a promising decision-support prototype for further translational research rather than an immediately deployable clinical screening system. While the demonstrates consistently strong diagnostic performance under benchmark validation, supporting its value as a proof-of-concept decision-support framework, future studies will involve external datasets and prospective clinical evaluation are essential to fully establish realworld screening applicability.

Informed Consent: Informed consent was no need Informed consent.

Conflict of Interest Statement: No benefits in any form have been received or will be received from a commercial party related directly or indirectly to the subject of this article. All authors declare no conflict of interest for this article.

Funding: No Funding support is provided for this paper publication.

Use of AI for Writing Assistance: No AI technologies utilized.

Author Contributions: Concept - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Design ? P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Materials - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Data collection and/or processing - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Data analysis and/or interpretation - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Literature search - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Writing - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.; Critical review - P.P., K.U., P.K., G.V.R., S.B.K., A.K., R.J.

Acknowledgments: The authors gratefully acknowledge all individuals and institutions who provided support and resources for the successful completion of this work.

Peer-review: Externally peer-reviewed.

Conclusion