Breast cancer is a severe disease. It happens when abnormal cells in the breast grow out of control (1). It is the second most common cancer found in women after skin cancer. In 2022, approximately 2.3 million women in the world were diagnosed with breast cancer, and approximately 670,000 died from it (2). Primarily, it affects women, but in some rare cases, men can also have breast cancer. According to abnormal cell growth, breast cancer is divided into two types: invasive and non-invasive (3).

The main risk factors of breast cancer include aging, genetic hereditary mutation, hormonal changes, lifestyle habits, and environmental factors (2). If not treated, the disease spreads to other organs and causes multiple organ failure and death (4). Early detection increases the survival rate of patients (1–3). The common symptoms of breast cancer are swelling or a lump in the breast, chest, or underarm, bloody discharge from the nipple, continuous pain in the breast, and a change in the size or shape of the breast (5).

Mammography is the most common screening method for breast cancer. It helps to find breast density, calcifications, structural problems, and tumor masses (6). Other diagnostic tests include magnetic resonance imaging (MRI), positron emission tomography (PET) scans, computed tomography (CT) scans, biopsy, gene testing, and estrogen or progesterone receptor tests. Treatment methods include surgery, chemotherapy, radiation therapy, hormone therapy, targeted therapy, and immunotherapy (4). Humans cannot control genetic factors, but a healthy lifestyle can reduce the chances of breast cancer (7). The preventive steps include less use of alcohol and smoking, keeping body weight normal, doing regular exercise, breastfeeding, and not using hormone therapy after menopause.

Even with advanced diagnostic technologies, some problems persist, including low imaging resolution, the inability to detect minor symptoms, high procedure costs, and slow diagnosis results (8). According to the National Institutes of Health (NIH), human error is the main reason for almost 96.3% of diagnostic failures (9). The use of machine learning (ML) can solve many of these problems. ML can learn patterns from a large amount of data, and it is helpful for disease prediction and early detection. Many studies show that ML can help doctors diagnose breast cancer correctly. By checking medical data, ML can improve medical decisions, reduce human mistakes, and make diagnoses and treatments faster (9).

Explainable artificial intelligence (XAI) is the method that makes ML results easy to understand for humans (10). XAI not only checks model predictions but also enhances the reliability, stability, and transparency of models, helping to identify and correct errors. Much research has been done on the use of ML and XAI for breast cancer detection and classification. One research used the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, which has 569 records with 32 features, and compared three ML algorithms for breast cancer detection. The k-Nearest Neighbors (kNN) algorithm achieved the highest accuracy of 95.9%, which demonstrates its strong performance in classification (11). Another research used the same dataset but focused on five main features and found that support vector classifier (SVC) gave better result with 93% accuracy (12). One more research used the Mendeley dataset, which has 400 Indonesian patient cases, in which 200 were patients with breast cancer, and found that the XGBoost model gave 85% accuracy. The SHAP (Shapley Additive Explanations) and CRISP MLQ frameworks were used to make ML results more understandable and to check the quality of the model (13).

An ensemble model that used support vector machine (SVM) and random forest (RF) algorithms got a very high accuracy of 99.99% (14). Another research, based on data from 500 patients at Dhaka Medical College Hospital in Bangladesh, used five ML algorithms and found that XGBoost yielded the best performance with 97% accuracy. The use of SHAP made the model clearer and increased trust in the results (15). One additional study utilized a pre-trained ResNet50 model to classify breast tumor images as benign or malignant, achieving 96.84% accuracy. The dataset of this study was taken from the Kaggle repository, and it has 7,909 breast tumor images of 82 patients collected through surgical open biopsy (16).

Despite the notable progress of ML and XAI techniques in breast cancer diagnosis, a critical methodological gap remains. Most existing ML/XAI approaches primarily focus on improving predictive accuracy and providing post-hoc explanations, while neglecting three key aspects required for reliable clinical deployment. First, current models rarely incorporate explicit uncertainty estimation, resulting in overconfident predictions that fail to indicate when model outputs may be unreliable. Second, widely used explanation techniques, such as SHAP and LIME, are primarily correlation-based and do not distinguish between causal clinical factors and spurious associations, thereby limiting their ability to identify proper diagnostic drivers and potential confounders. Third, fairness assessment across demographic or temporal patient subgroups is often overlooked, raising concerns about biased or inconsistent performance in real-world settings. Importantly, these limitations are typically addressed in isolation, if at all, and no unified framework jointly integrates uncertainty awareness, causal reasoning, and fairness analysis within a single diagnostic pipeline. This gap restricts the clinical trustworthiness, transparency, and ethical applicability of existing breast cancer prediction systems.

To address these problems, this research presents a new Uncertainty-Aware Causal Explainable Ensemble Framework for breast cancer prediction. The proposed method integrates uncertainty estimation, causal feature analysis, and multimodal explainability into a single ensemble model. By giving predictions with information of when the model is not sure, the framework improves the reliability, understanding, and fairness of diagnosis, and makes a base for trusted clinical decision support in real healthcare use.

The experimental testing was carried out by using two datasets: one is the UCTH Breast Cancer Clinical Dataset (17) and the other is the Breast Cancer Wisconsin Diagnostic Dataset (18). The use of these two datasets gives both clinical and morphological validation of the proposed framework.

This research presents an Uncertainty-Aware Causal Explainable Ensemble Framework for breast cancer prediction, which includes uncertainty calculation causal feature analysis multimodal explainability and fairness checking in one reproducible structure. Figure 5 shows the complete workflow of the proposed framework, which includes the step-by-step process of preprocessing feature engineering model development uncertainty estimation causal analysis and fairness evaluation.

The methodology started with data preprocessing in which the clinical breast cancer dataset was checked for consistency and completeness. Missing values were filled by using statistical estimators; categorical variables were changed by one-hot encoding and numerical features were standardized for balanced scaling. To include time-based changes, a Temporal Feature Engineering module created extra features like diagnosis era years since first record and age–era interaction that help the model learn from long-term diagnostic variations and time-based clinical trends.

An Uncertainty Aware Ensemble Model was developed, which used three different classifiers—LightGBM, RF, and gradient boosting—to take advantage of different learning biases. Each classifier gave probability outputs, and these outputs were combined to calculate the mean predictive probability of all models. The standard deviation of these probabilities was used to measure epistemic uncertainty, which shows the disagreement between ensemble models. Predictions with uncertainty higher than threshold (τ = 0.15) were marked as uncertain to produce qualified predictions that focus on reliability more than coverage.

For the ensemble, we chose LightGBM, RF, and gradient boosting because they provide complementary strengths and work well on small to medium clinical tabular datasets. RF reduces variance, gradient boosting learns error correcting patterns, and LightGBM offers fast histogram-based learning with strong support for mixed feature types. These models also produce stable probability outputs, which help with uncertainty estimation and SHAP-based interpretability. We tested other models like XGBoost and simple neural networks, but XGBoost showed higher variance and weaker uncertainty calibration, while neural networks required more data and tuning. The three selected models offered the best balance of accuracy, stability, interpretability, and efficiency, making them suitable for clinical deployment.

To find statistically important and causally related variables, a Causal Feature Analyzer was used, which measured the relation between input features and diagnostic results by using Cramer’s V for categorical features and point biserial correlation for continuous features. This analysis showed that Invasive Nodes (Inv Nodes) and Tumor Size are the main clinical factors that affect malignancy. A Robust Multimodal Explainer was also added to improve interpretability and clarity by combining information from SHAP permutation importance and model feature importance. When SHAP calculation was not possible, a permutation-based backup was used to keep interpretability working and reliable for all samples.

Model reliability and generalization were tested by using bootstrap-based confidence interval estimation with 500 to 1,000 iterations for performance measures like accuracy, precision, recall, F1 score, and AUC. A Fairness Assessment Module checked predictive equality for different demographic groups like age and diagnostic era and measured the difference in performance to make sure that the model works fairly. Overall, this framework combines ensemble learning with causal and explainable AI ideas to achieve prediction performance interpretability and fairness suitable for clinical use.

The proposed architecture is a modular uncertainty-aware and causally explainable ensemble designed to provide diagnostic accuracy and clinical trust. As shown in Figure 6, the framework has five connected layers: (1) Data preprocessing and feature engineering, (2) Base ensemble learning, (3) Uncertainty quantification layer, (4) Causal explainability layer, and (5) Statistical fairness validation. Each part gives its own function to make a trusted AI-based breast cancer prediction system. The system begins with a structured dataset D:

where xi∈ℝd is the ith input vector containing demographic, pathological, and clinical features, and yi∈{0,1} represents the ground truth label (0 = benign, 1 = malignant). Data normalization is applied to maintain statistical consistency:

where µ and σ denote the mean and standard deviation of each feature, respectively.

1. Temporal feature augmentation: To model diagnostic changes, a temporal encoder creates time-based attributes that represent diagnosis period recency and the relation between age and diagnostic period. These features help the model to learn generational and time-dependent changes in breast cancer characteristics.

2. Heterogeneous ensemble learning: The model ensemble consists of three classifiers—LightGBM, RF, and gradient boosting—chosen for their complementary learning biases and robustness to clinical data heterogeneity. Each base learner M_k outputs a class probability pk(y=1|xi) for the ith sample. The ensemble mean probability and epistemic uncertainty are computed as:

where K = 3 denotes the number of models. The final prediction rule incorporates an uncertainty threshold τ:

This mechanism ensures that predictions marked as uncertain are deferred for clinical review, safeguarding against overconfident misclassifications.

3. Causal feature analysis: To enhance interpretability, causal inference techniques identify confounding variables that may distort model learning. Categorical relationships are quantified using Cramer’s V:

while continuous–binary associations are captured by the point-biserial coefficient:

Features with |r_pb|> 0.1 or V > 0.1 are considered potential confounders, reflecting significant diagnostic influence.

4. Multimodal explainability integration: This layer fuses diverse interpretability techniques SHAP, permutation importance, and model-native feature importance to deliver consistent, clinician-understandable insights. The SHAP contribution for each feature x_j is expressed as:

and permutation-based relevance as:

The consensus importance metric integrates all three views:

5. Statistical and fairness validation: The reliability of the ensemble is established using bootstrap-based confidence intervals:

where M^∗ represents bootstrap-resampled metric estimates. Fairness is evaluated by comparing subgroup performance:

ensuring equitable outcomes across demographic groups such as age and diagnostic era.

The proposed architecture makes a closed-loop diagnostic intelligent system where uncertainty interpretability and fairness are the main design principles. Each module helps to produce clinically meaningful and confidence-aware predictions that follow the needs of trusted AI- and evidence-based medical decision-making.

The integration of these parts allows the framework to work as a transparent and self-evaluating diagnostic system. By combining uncertainty calculation causal reasoning and explainability fusion, the model not only predicts breast cancer outcomes but also explains its decisions in a clinically understandable way. This complete design makes sure that each diagnostic result has measurable confidence and fairness indicators that build trust reproducibility and accountability in AI-based clinical decision-making. The modular structure also allows easy connection with different clinical datasets and changing diagnostic protocols without retraining the whole model. The ability of the architecture to measure uncertainty while keeping explainability supports clinical triage model validation and decision support. In the end, the proposed system gives a base for future use of reliable human-aligned and regulation-ready AI tools in medical diagnostics. The proposed uncertainty-aware causal-explainable ensemble is described in Algorithm 1 for training and Algorithm 2 for deployment-time inference with uncertainty estimation and explanation.

The performance and explainability of the proposed Uncertainty-Aware Causal Explainable Ensemble Framework were fully tested in many aspects like prediction strength uncertainty calculation causal interpretability explainability fairness and clinical importance. The testing used two datasets together to make sure of clinical realism and experimental generalization.

The first dataset, the Breast Cancer Clinical Dataset (see Table 1), has 213 patient records that include demographic pathological and time-related variables. After preprocessing and adding time-based features, the dataset has 20 engineered predictors. The data were divided in a 70:30 ratio with 149 training and 64 testing samples and a balanced class distribution (benign = 120 and malignant = 93). This dataset gave a real-world clinical base for testing the model explainability and uncertainty-aware ability on mixed tabular data.

To validate the robustness and scalability of the proposed architecture, experiments were also conducted on the publicly available Breast Cancer Wisconsin (Diagnostic) dataset (see Table 2), which consists of 569 biopsy samples derived from digitized fine-needle aspirate (FNA) images of breast tissue. Each instance contains 30 numerical features representing morphological characteristics of cell nuclei—computed as mean, standard error, and worst values spanning properties such as radius, texture, area, and concavity. The benchmark dataset enabled comparative evaluation against existing ML approaches, demonstrating the framework’s ability to generalize beyond small-scale clinical data while maintaining explainability and fairness.

To test the strength and scalability of the proposed architecture, experiments were also done on the public Breast Cancer Wisconsin (Diagnostic) dataset (see Table 2), which has 569 biopsy samples taken from digital FNA images of breast tissue. Each record has 30 numerical features that show the shape and structure of cell nuclei calculated as mean standard error and worst values for properties like radius texture area and concavity. This benchmark dataset helped to compare the framework with existing ML methods and showed that the framework can work beyond small clinical data while keeping explainability and fairness.

Together, these two datasets provided complementary understanding: the clinical dataset captured time-based and demographic changes important for causal analysis, and the benchmark dataset focused on detailed morphological features. Both datasets together allowed complete testing of diagnostic accuracy and model trust in the medical AI field.

While the framework uses Cramer’s V and point biserial correlation to identify important clinical confounders, the current causal analysis is still correlation based. These measures help show which features have strong relationships with malignancy, but they do not provide true directional causality or model counterfactual situations. In future work, we plan to improve the causal module by adding causal discovery methods such as PC, FCI, or NOTEARS to learn causal graphs directly from clinical data. We also aim to include structural causal models and counterfactual reasoning so that the system can simulate “what if” scenarios, such as how the malignancy risk changes if a specific clinical variable is modified. These additions will move the analysis beyond correlations and provide stronger causal explanations, improving clinical interpretability and supporting better decision-making in precision medicine.

Our uncertainty quantification approach is based on ensemble disagreement, which captures epistemic but not aleatoric uncertainty. While this is suitable for tabular clinical datasets where epistemic uncertainty dominates, future work could incorporate noise-aware models (e.g., Bayesian neural networks or Monte Carlo dropout) to account for data variability, particularly in imaging or temporal health records. Nevertheless, the current framework’s ability to identify low-confidence cases provides a clinically useful safeguard against overconfident predictions.

While the proposed framework demonstrates strong and consistent performance on retrospective datasets, its clinical readiness must be validated through external multicenter studies and prospective trials. Future work will focus on evaluating the model in diverse healthcare settings, across varied demographic and clinical populations, and in real-time diagnostic workflows to confirm generalizability, robustness, and practical utility.

Recent studies on breast cancer prediction have increasingly integrated ML with XAI to improve diagnostic reliability and transparency. Islam et al. (20) evaluated several traditional classifiers, including SVM, RF, Logistic Regression (LR), Gradient Boosting Classifier (GBC), K-Nearest Neighbors (KNN), XGBoost, and Decision Tree Classifier (DTC), on the Breast Cancer Wisconsin (Diagnostic) dataset. Their SVM model achieved an accuracy of 98.25% (AUC ≈ 0.98) and an F1 score of 0.99, using SHAP and LIME for interpretability. While the study demonstrated strong diagnostic performance, it lacked mechanisms for uncertainty estimation, causal feature discovery, and fairness evaluation critical components for clinical trustworthiness.

Similarly, Arravalli et al. (19) introduced a stacking ensemble combining nine classifiers and five interpretability methods (SHAP, LIME, ELI5, QLattice, and Anchor). Their framework achieved an AUC of 0.96 on the UCTH dataset, emphasizing explainability but without integrating uncertainty or fairness assessment.

In comparison, the proposed Uncertainty-Aware Causal Explainable Ensemble Framework improves the existing models by adding uncertainty calculation causal analysis and fairness checking directly in the prediction process. When tested on the WDBC dataset, the framework achieved an AUC of 0.996, an F1 score of 0.984, and an accuracy of 0.997 under normal conditions with a certain-only accuracy of 0.987 and a precision of 1.000. In the UCTH dataset, it achieved an AUC of 0.95 and a certain accuracy of 0.93 with no false positives. These results show that the proposed model gives a state-of-the-art performance with more transparency stability and clinical reliability and solves the main problems of previous XAI-based models. Table 9 compares the proposed framework with existing state-of-the-art breast cancer prediction models, highlighting its superior performance and added integration of uncertainty, causal analysis, and fairness evaluation.

To check the separate effect of each part in the proposed Uncertainty-Aware Causal Explainable Ensemble Framework, an ablation study was done by turning off specific modules one by one and observing the changes in prediction performance, uncertainty calibration, and interpretability stability. The base setup Base Ensemble used only the three classifiers LightGBM, RF, and gradient boosting without uncertainty calculation, causal analysis, or fairness checking. Each next setup added one module to see its individual effect.

The results shown in Table 10 indicate that uncertainty calculation clearly improved diagnostic reliability by detecting 10.9% uncertain cases, which increased precision by 9.6% for certain predictions. Adding the Causal Feature Analyzer improved feature explainability and confounder detection, making feature importance match with verified clinical factors like Inv Nodes, Tumor Size, and Age. Adding the Multimodal Explainability layer made interpretability more stable across SHAP permutation and model importance measures and produced stronger and more repeatable insights. The Fairness Assessment module reduced subgroup difference by lowering accuracy gap by 0.12 and ensured fair generalization for age and diagnostic era groups.

Overall, each module added specific improvement in performance transparency and ethical compliance, which confirms the framework design and its complete alignment with trusted clinical AI principles.

These results confirm that each part of the architecture, including uncertainty calculation, causal analysis, explainability, and fairness, together increase the model’s diagnostic credibility. The final model reaches a balanced point between prediction strength and interpretability and performs better than the base ensemble in both clinical transparency and reliability.

To check the practical use of the proposed framework in real clinical environments, we evaluated its computational needs and possible integration methods. The model has a lightweight inference design because the ensemble uses tree-based algorithms (LightGBM, RF, and gradient boosting), which are much faster than deep learning models. Testing using a normal workstation (Intel i9 CPU, 64 GB RAM) revealed that the average inference time for one patient sample was less than 8 ms, which shows that the system can run in real time inside hospital workflows without a GPU. For deployment, the framework can be added as a modular decision-support tool that connects with existing hospital systems such as PACS, EHR, or LIS. Since the model works on structured clinical and morphological features, it can be deployed through a REST API or as an on-premise microservice inside the hospital IT setup. The uncertainty-aware outputs give clear triage information, helping clinicians automatically flag high-risk patients and manually check uncertain cases, which improves workflow safety and reduces diagnostic delay. Because of its low computational cost, clear explainability, uncertainty estimation, and easy connection with existing data pipelines, the proposed method is suitable for large-scale use in real clinical environments.

In conclusion, this study introduced a reliable prediction framework that combines ensemble learning with uncertainty calculation to improve performance and explainability in clinical analysis. The proposed approach combines multiple base models through ensemble integration and uses uncertainty estimation to detect and remove low confidence predictions, which increases the reliability of clinical decision-making. Experiments on two benchmark datasets showed strong and consistent results. On the Breast Cancer Wisconsin Diagnostic dataset, the framework achieved an AUC of 0.99, an accuracy of 0.98, and an F1 score of 0.98. On the UCTH clinical dataset, the model achieved an AUC of 0.97, an accuracy of 0.95, and an F1 score of 0.94 with a perfect precision (1.000) for certain predictions and no false positives. It is important to note that these results are based on retrospective datasets of moderate size and that perfect precision was achieved only after filtering uncertain predictions; prospective validation in larger, real-world clinical cohorts is needed to confirm generalizability. The causal feature analysis confirmed the clinical relevance of key predictors like lymph node involvement, metastasis, and tumor morphology and the fairness evaluation showed stable performance across age groups with a small difference (ΔF1 = 0.200). These results confirm that the framework provides accurate explainable and fair breast cancer predictions across different clinical datasets. Although this study focuses on breast cancer, the proposed Uncertainty-Aware Causal Explainable Ensemble Framework is not limited to this disease. The main components such as ensemble learning, epistemic uncertainty estimation, causal feature analysis, and multimodal explainability are general and can be applied to other cancer types including lung, cervical, colorectal, or prostate cancer. These cancers also use diverse clinical and morphological features, where uncertainty handling and clear interpretability are important. Because of this, the framework has strong potential to generalize across different cancers, with only dataset-specific preprocessing and clinical validation needed. Future research will focus on increasing dataset diversity, improving uncertainty calibration, and testing generalization on wider healthcare environments to support reliable and ethically aligned AI clinical decision systems.