Detecting and Mitigating Data Leakage in Breast Cancer Prediction Using an Explainable Machine Learning Framework with XGBoost and SHAP

Breast cancer is the most common cancer in women worldwide and is still one of the main causes of cancer death. According to the updated cancer statistics for 2020, there were more than 2.3 million new cases and approximately 685,500 deaths from breast cancer globally [1], which evidences the urgent need for powerful diagnostic and prognostic approaches that enrich clinical decision-making and personalize therapy.

Machine learning approaches have recently been applied to breast cancer in diagnosis, prognosis, and treatment-response estimation areas [2-4]. Several algorithms, including logistic regression, random forest (RF), support vector machines (SVM), and gradient boosting algorithms such as Extreme Gradient Boosting (XGBoost), achieved promising predictive results on clinical-pathological-imaging-derived datasets [2, 4, 5]. Due to their ability to capture complex, non-linear relationships in the data as well as handle heterogeneous high-dimensional features in a natural manner with an empowered framework for a fully data-driven clinical decision-making process, these approaches have become increasingly popular tools among researchers in oncology.

However, despite these advances, increasing criticism has recently been raised over the methodological quality, transparency, and clinical applicability of machine learning-based healthcare studies [6-8]. High predictive performance of a machine learning algorithm does not equal clinical utility or generalizability to real-world patient groups. However, numerous medical artificial intelligence (AI) systems were found to have reproducibility problems related to inappropriate validation procedures, methodological flaws, and hidden biases that might artificially increase predictive performance [6, 8]. This has increased the need for a framework that reflects reliable, clinically credible AI technology in health care [9].

Data leakage, which is a methodological issue [6, 10], presents one of the most serious and unrecognized threats to trustworthy predictive modelling within clinical machine learning. Data leakage is defined as the inclusion of information that should not have been available when using data that would not be available at prediction time in the final prediction-making process, for example during training or testing of a model [10]. This is an especially problematic issue in a clinical scenario where patient records typically contain longitudinal follow-up variables, treatment outcomes or survival indicators, or post-diagnostic measurements that can act as indirect encoders for the target label [8, 11]. As a result, machine learning models may instead learn future outcome information rather than clinically meaningful prognostic patterns, ultimately leading to unrealistic predictive performance and poor generalizability in the medical domain [6, 10].

Machine learning research has identified various forms of data leakage, including target leakage, temporal leakage, and preprocessing leakage [10, 12]. Especially in clinical prediction tasks, target leakage is problematic because variables that are linked to the outcome can be included when the model is created accidentally. Temporal leakage appears when future clinical information comes into the training dataset, while preprocessing (or pipeline) leakage takes its place when average methods of scaling, normalization, imputation, or feature encoding are applied prior to suitable separation between train and test datasets [10, 12]. The most prevalent types of leakage found in clinical machine learning applications are summarized in Table 1.

Table 1. Common types of data leakage in clinical machine learning

Explainable artificial intelligence (XAI) methods help study the transparency and interpretability of these models, which has become more important for clinical applications of healthcare machine learning [13, 14]. Of the many existing approaches, SHapley Additive exPlanations (SHAP) have become one of the most popular model-agnostic interpretive techniques for machine learning-based predictions [15]. SHAP offers global and local interpretations by estimating the contribution of each feature to model predictions. In oncology research, SHAP has mainly been applied in the context of feature importance and model interpretation [13]. However, its applications in auditing methods for implicit data leakage and unreliable predictive patterns are limited.

Newer guidelines for clinical AI, such as Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD)-AI and trustworthy AI frameworks [9, 16], highlight the importance of validation that avoids relevant leakage and transparent reporting. These approaches highlight the demand for methodological rigor, transparency of analytic techniques, and reproducibility, as well as clinically relevant assessment methods that are crucial to enable and allow efficient (and safe) translation of machine learning systems toward real-world health care models. Yet in that field, while the knowledge about reproducibility and reliability of AI output is increasing, there have been no definite studies of data leakage as a principal methodological focus in breast cancer prediction. Most existing studies focus on optimizing predictive performance and are not systematically validated to determine whether the learned patterns are clinically meaningful or temporally appropriate.

Motivated by these challenges, this work examines where data leakage occurs and how to detect and mitigate it with a specific focus on clinical machine learning models for breast cancer prediction. Unlike traditional predictive works that target solely finding the model that maximizes classification accuracy, the present work emphasizes methodological soundness, leakage-aware evaluation, and interpretability-driven auditing. To strive for this goal, this study proposes a Leakage-Aware Interpretability Framework (LAIF), which combines machine learning with SHAP-based model explainability, feature auditing, and domain knowledge to systematically identify potential leakage-prone variables and take measures.

We evaluated the proposed framework on two heterogeneous breast cancer datasets to study the impact of information leakage when predicting patient outcomes in clinical practice by simulating realistic clinical conditions with respect to predictive performance, feature importance, and model interpretability. The proposed framework adds a practical and reproducible dimension to incorporating explainability into the validation process to improve the reliability, transparency, and clinically relevant trust in an increasingly common layer of AI for oncology.

3.1 Datasets

In this study, we employed two publicly available breast cancer datasets that differ in nature to test the model performance under heterogeneous settings and to examine how data leakage may influence predictive models.

3.1.1 Mendeley dataset

The primary dataset for this study is from the Mendeley Data repository, which contains a large breast cancer cohort with demographic, pathological, molecular, treatment-related, and outcome-associated variables. Following data acquisition, the dataset consisted of 15,054 patients and 34 variables: a systematic aggregation of routine clinical information shared in most oncology studies. It contains patient characteristics, tumor descriptors, receptor status information, treatment indicators, as well as molecular subtypes and survival outcomes [19].

Relapse-free status was selected as the prediction target for binary prediction. In total, we included 8,976 patients labelled as Not Recurred and 6,072 patients labelled as Recurrence, a very imbalanced class dataset. This distribution is similar to that observed in clinical practice and requires an evaluation metric beyond conventional accuracy [20, 21].

Missing values: Preprocessing-stage median and mode imputation for numerical/categorical features was used to address missing observations in some variables. Since patients within a dataset remain unchanged and lose less information, these imputation methods are well-utilized in many clinical machine learning studies [22, 23].

One unique aspect of this dataset is that there are several follow-up and outcome variables, including overall survival status, overall survival (months), relapse-free status (months), and patient’s vital status. However, predictive variables such as these are not present at the time of initial clinical decision-making, when intervention is to be guided. Therefore, using them as features when training a model can introduce target leakage and lead to overly optimistic performance estimation [2, 18].

The dataset is suitable for the purpose of demonstrating leakage detection and mitigation strategies within clinical machine learning as it contains several baseline clinical variables and information that is dependent on the outcome. This property of the dataset makes it especially appropriate for testing whether interpretability techniques can help not only detect potential leakage-inducing features but also increase the trustworthiness of predictive models.

A heterogeneous set of clinical, pathological, molecular, treatment-related, and outcome-associated variables from a large breast cancer cohort is summarized in Table 2, representing the Mendeley dataset. Furthermore, the existence of survival and follow-up features introduces a high-risk situation for target leakage, making it ideal in many respects to evaluate leakage detection and degradation methods within clinical machine learning.

A heterogeneous set of clinical, pathological, molecular, treatment-related, and outcome-associated variables from a large breast cancer cohort is summarized in Table 2, representing the Mendeley dataset. Furthermore, the existence of survival and follow-up features introduces a high-risk situation for target leakage, making it ideal in many respects to gauge both leak detection algorithms as well as degrading methods within clinical machine learning.

Table 2. Summary characteristics of the Mendeley breast cancer dataset used in this study

3.1.2 Kaggle dataset

The Breast Cancer Wisconsin Diagnostic dataset from Kaggle was also used for benchmarking. The dataset contains 569 samples that are characterized using a total of 30 numerical features that have been computed from the digitized Fine Needle Aspiration (FNA) images [24].

This variable is known to be the target variable in tumor diagnosis, i.e., either malignant or benign. It is different from the Mendeley dataset, which provides not only baseline diagnostic measurements but also follow-up information, survival outcome, and post-treatment variables; whereas the Kaggle dataset contains only baseline information. Accordingly, the exercise with limited risk of target leakage from one dataset to another serves as a suitable reference dataset to validate our proposed modeling pipeline in conditions close to leakage-free [18].

Utilizing both datasets together allows for a synergistic evaluation approach, with the clean Kaggle benchmark allowing for initial testing of the model, and the Mendeley dataset providing real-world clinical challenges for external validation characterized by heterogeneous distributions across features, outcome-dependent variables, and high leakage risk [6, 10].

3.2 Data preprocessing

Both datasets underwent consistent preprocessing pipelines to ensure reproducibility and minimize methodological bias.

For numerical features, median imputation was used to fill in missing values, and the most common category for categorical variables [22, 23]. Only records with excessive missing values were filtered out from the analysis. Before transformation, features were divided into numerical and categorical types.

Z-score normalization was applied to numerical features. Crucially, scaling parameters were calculated on the training data only and then applied to the test set to prevent leakage of preprocessing [10, 12].

One-hot encoding was used to encode categorical variables in the clinical dataset. Finally, all preprocessing steps were done inside a pipeline to make sure that transformations were only learned using the training data [12].

We used stratified sampling to split the datasets into training and testing subsets. To achieve an unbiased evaluation, the test set remained fully separated from all preprocessing and model training pipelines [21].

Prior to developing the models, class distribution was analyzed. The Mendeley dataset had mild class imbalance, where there were 8,976 non-recurred and 6,072 recurred cases. But without applying any synthetic oversampling techniques. However, model performance was evaluated using relevant and robust learning evaluation metrics, respectively: precision, recall, F1-score, and receiver operating characteristic-area under the curve (ROC-AUC), which are also more appropriate than accuracy for evaluating most moderately imbalanced classification models [20, 21].

3.3 Proposed Leakage-Aware Interpretability Framework

We introduce LAIF that seamlessly blends machine learning, interpretability, and feature auditing to systematically diagnose and prevent data leakage. The data preprocessing, predictive modeling, SHAP-based interpretation, domain-expert assessment, and leakage auditing are integrated into a unified validation workflow that is proposed to verify the model interpretability [2, 13, 18]. Figure 1 shows the general architecture of this framework.

Figure 1. Proposed Leakage-Aware Interpretability Framework (LAIF) integrating SHapley Additive exPlanations (SHAP) analysis and domain knowledge for systematic leakage detection, feature auditing, model retraining, and leakage-free evaluation

Data preprocessing is the first step in the workflow, where different stages like handling missing values, encoding features, scaling features, and balancing classes are considered. We then organise the data into a training and test split to create an initial XGBoost classification model. The performance of models is evaluated using standard classification metrics like accuracy, precision, recall, F1-score, and ROC-AUC.

To enhance transparency and credibility, global and local interpretations of model behavior are presented using SHAP analysis. The development of a machine learning workflow usually involves model explainability being used as a post-exploitation process, but in the proposed framework, SHAP is integrated into the model validation. SHAP-based feature auditing is combined with domain-expert assessment of where information leakage [12, 13] might enter the data-generating process.

Features that were identified as being prone to leakage are forcibly audited and removed before the model is recreated. After removing leakage features, we re-trained the model using only clinically available baseline variables and then re-evaluated it to obtain leakage-free estimates of performance. This avoids spurious overfitting, thereby strengthening model robustness, while optimizing clinical generalizability of machine learning predictions [2, 4].

We present the LAIF framework, which enables a synergistic combination of explainability, domain knowledge, and leakage auditing to improve model transparency, bolster methodological rigor, and facilitate application development of clinically reliable machine learning systems.

3.4 Feature encoding

One-hot encoding was used to encode the categorical features of the clinical dataset to better capture their nominal nature [12].

The imaging dataset was not encoded since it contains only numerical features.

Over-and under-sampling methods were applied; this was because the pipeline works only on training data to avoid leakage, so it cannot be trained or fit on unseen data, but due to its design it does not suffer from any leakage [10, 12].

To facilitate SHAP interpretability analysis, feature names are kept across the entire pipeline, so that all encoded features can easily be mapped back to their clinical meaning [9, 18].

3.5 Model development

Of the several classification algorithms tested, XGBoost was selected as our main classifier because it was selected from several classifiers due to its strong performance on structured clinical data with heterogeneous features and is used generally for structured clinical data with continuous and heterogeneous features [2, 4].

We trained the model using a gradient boosting framework, which sequentially fits decision trees to minimize prediction error [3]. Hyperparameters, including the number of trees, maximum tree depth, and learning rate, were tuned using grid search with cross-validation on training data.

We applied regularization to avoid overfitting and gain better generalization, using techniques such as subsampling and depth constraints [12, 25]. Model selection occurred using only the ROC-AUC due to its robustness in evaluation compared to accuracy for moderately imbalanced classification problems [20, 21].

Other baseline models, including logistic regression and RF, were also tested for comparison.

3.6 Explainability analysis (SHapley Additive exPlanations)

Model predictions were interpreted using SHAP [15], and feature importance was analyzed. SHAP generates global and local explanations by approximating feature contributions to model output.

Global SHAP summary plots made it possible to see the most important features, and dependence plots allowed for analysis of how feature values affect predictions [15, 18].

SHAP was also used as a diagnostic tool to detect possible data leakage, in addition to its use as an interpretation tool. Any features with unusually strong influence were marked for follow-up investigation, notably outcomes-related features [6, 12].

3.7 Data leakage detection and mitigation

A systematic approach to identify and prevent data leakage was used. Initially, features were screened for clinical plausibility to identify features absent at prediction time. Second, SHAP analysis can be performed as a feature with excessively high importance [6, 12].

Early on, using all features gave very high predictive power, meaning there was probably unwanted target leakage. Outcome-related variables dominated model predictions, and this was further established with SHAP analysis [4].

We then removed these features and retrained the model with a clean dataset. We recreated the preprocessing pipeline, without using any information from the test set during training. This allowed pre- and post-cleaning model performance under leakage-aware conditions [12] to be compared without the need for environmental cleaning.

To mitigate potential leakage from unknown variables, a systematic review was done of all identified variables. Variables were categorized into those that would directly or indirectly cause leakage and eliminated if they leaked information unavailable at the planned prediction time [10, 26]. A summary of the leakage-prone variables and reasons for exclusion are presented in Table 3.

Some possible leakage features were detected by a two-stage auditing process. The first step was to study the SHAP summary plots, which enabled us to know which variables were especially emphasizing the prediction. The second step involved assessing the clinical focus of each top feature using domain knowledge to ascertain whether the associated information would be discernible at the time point intended for prediction. Those features that showed both high SHAP importance and contained information from either a post-outcome, follow-up, or survival-related point in time were labelled as leakage-prone variables and excluded from the final model [6, 10, 13]. This process made the leakage detection procedure more reproducible and transparent.

Performance dropped to more realistic but still robust levels after cleaning, suggesting that the model was identifying generalizable patterns in the data rather than leaking information [6, 26].

Table 3. Leakage-prone variables identified and removed from the Mendeley dataset [6, 10, 18]

3.8 Evaluation metrics

Model performance was measured in terms of multiple classification metrics including accuracy, precision, recall, F1-score, and ROC-AUC.

While accuracy was reported, the focus was on recall and F1-score due to their clinical relevance in reducing false-negative predictions. This study used ROC-AUC as the main evaluation metric, which is a robust metric to evaluate moderately imbalanced classification [20, 21].

Moreover, precision–recall (PR) curves were also analyzed to offer an alternative assessment in the context of class imbalance [21].

We used stratified cross-validation for model training and evaluation to prevent information leakage between the training and testing data [10, 12]. The reproducibility of experimental setups was controlled by the use of fixed random seeds in stochastic methods [6, 12], and all experiments were implemented over fully reproducible pipelines.

We show that data leakage encodes highly misleading signals which may distort clinical machine-learning models, stressing the importance of interpretability-driven validation to ultimately yield reliable and clinically relevant results. The results indicate that, based on the assessment of model behavior across two heterogeneous datasets, our proposed XGBoost framework is relatively robust to distributional shifts in training data, outcome-dependent feature contamination poses considerable risk.

The testing results on the Kaggle dataset show a strong leakage-free baseline. High predictive power is obtained by using strictly defined imaging-derived features, and this shows that the proposed preprocessing and modelling pipeline can capture informative signal in leakage-free conditions. This trend is also corroborated by the corresponding SHAP analysis, establishing that the model utilized clinically justified imaging morphology features such as tumor size and texture-related features. When the importance conferred by the data and our understanding of its process are aligned, we gain confidence in both the validity and interpretability of our model.

Compared to the initial experiments on the Mendeley dataset, which clearly overfitted and reached unprecedented predictions, with predictions very close to 1s (not realistic for a real-world clinical prediction task). That level of performance showed that leakage is occurring, rather than indicating a more capable model. Incorporation of domain knowledge and SHAP-based explanations revealed outcome-dependent variables as explanatory factors; survival status and relapse-driving characteristics proved to be dominant drivers of model predictions. This would be a clear scenario of target leakage, and the worst methodological pitfall imaginable, even if you do not have these variables when making predictions (which, of course, represents one of the canons on which machine learning methodology relies).

The model performance is therefore more reasonable after eliminating features that are related to leakage. The importance of the reduction is not due to model degradation; rather, it shows that shortcut learning has been removed, and clinically valid prediction behavior has been adopted. Importantly, the model retained reasonable discrimination, suggesting that baseline clinical characteristics carry some meaningful prognostic information. This result lends additional evidence to the resource-inefficient premise that, for many contexts with suitable validation processes, good prediction is achievable with outcome-derived variables.

From a methodological perspective, this work shows how SHAP interpretation methods can aid in both explaining model predictions and conducting audits of dataset integrity. Detection of leakage even in complex clinical datasets cannot be fully addressed by traditional evaluation metrics. This work presents a practical and reproducible methodology, primarily because its validation can be incorporated into existing explainability pipelines to uncover and mitigate hidden sources of bias.

On a clinical front, the implications highlight how deploying models trained on tainted data can be perilous. One consequence of such leakage is that models with poor generalization ability might appear to perform well in development but would subsequently underperform in real-world settings, leading to inappropriate conclusions about risk and clinical decisions. To mitigate this, we propose a leakage-aware framework to ensure that predictions are only made based on features that can be observed at the intended point of use, making them more reliable and leading to safer clinical decision-making.

This work integrates SHAP as a central model validation step within the modeling pipeline, in contrast to many previous studies that considered explainability a post hoc visualization step. This method allows for systematic leakage identification and better understanding of model behavior. While breast cancer datasets are used in this current analysis, the framework proposed here is applicable to other clinical disciplines that involve longitudinal or outcome-rich data.

In summary, our work provides a practical and reproducible approach to robustness-aware machine learning with interpretability as part of the model validation process instead of providing an optional post hoc analysis.

This study also illustrates how data leakage can invalidate a clinical machine learning model and the importance of validation that is driven by interpretability for reliable performance assessment. Using an XGBoost-based framework, we demonstrate that near-perfect predictive performance of survival can arise from the unintended inclusion of outcome-dependent features rather than true prognostic learning when evaluated on two heterogeneous breast cancer datasets.

Using systematic guided leakage detection coupled with SHAP-based analyses, we identified and removed outcome-related variables from the feature list and achieved improved leak-free performance characteristics that are measured in ways that better approximate clinical use. After removing leakage, predictive metrics were lowered, and the model kept high discriminative power, indicating that baseline clinical features encode real predictive signal.

This study demonstrates that it shows that explainability methods can not only be used as model interpretability tools, but also as effective ways to identify methodological weaknesses such as data leakage.

The proposed framework provides guidance on how to develop clinically relevant prediction models by ensuring that performance reflects data available at the time of prediction.

This work emphasizes the necessity of leakage-aware evaluation and interpretability-centered validation for clinical machine learning, leading to more reliable and clinically deployable AI systems.

The data used to support the findings of this study are available from the corresponding author upon request.