Aug 2026· AppliedMath· Vol 6, pp. 134· 0 citations· 21 references
TL;DR
This study systematically compared statistical and machine learning-based imputation methods using two publicly available breast cancer datasets representing complementary clinical settings to highlight the importance of considering dataset characteristics, missing-data mechanisms, and the intended analytical objective when selecting imputation methods.
Abstract
Missing data remain a major challenge in breast cancer research because they can introduce bias, reduce statistical efficiency, and compromise the performance of predictive models. Although numerous imputation techniques have been proposed, their comparative performance under different missing-data mechanisms and their impact on downstream classification remain inadequately understood. This study systematically compared statistical and machine learning-based imputation methods using two publicly available breast cancer datasets representing complementary clinical settings. The methods were evaluated under simulated Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR) mechanisms using both reconstruction accuracy and downstream classification performance. The results showed that no single imputation method consistently achieved the best performance across both datasets. Regularized regression and machine learning-based methods generally outperformed conventional statistical approaches, although the optimal method depended on the characteristics of the dataset. Furthermore, the best-performing imputation methods preserved downstream classification performance despite the introduction of missing data, demonstrating that reconstruction accuracy alone is insufficient for selecting imputation strategies intended for predictive modelling. Overall, the findings highlight the importance of considering dataset characteristics, missing-data mechanisms, and the intended analytical objective when selecting imputation methods. The proposed evaluation framework provides a robust approach for assessing missing-data handling strategies in breast cancer prediction studies and other biomedical machine learning applications.
Background and Objective: Reliable medical research depends on data integrity, yet clinical datasets often contain real and systematically missing values. This study aimed to develop a robust, clinically realistic imputation framework for a raw Multiple Sclerosis (MS) dataset affected by real-world missingness. Methods: We propose an innovative data-driven approach called the Sequential Multiple Imputation Bootstrapping (SMIB) model, which orders imputation based on feature correlation and incorporates bootstrapping to enhance stability and generalizability. Relevant features were identified using RF importance and Mutual Information scores and imputed using a hybrid machine learning framework that combines Random Forest (RF), Multilayer Perceptron (MLP), k-Nearest Neighbors (kNN), and a multiple imputation (MI) algorithm. The proposed method was evaluated using 15-fold cross-validation and a masking-based evaluation strategy. Model performance was assessed using accuracy, precision, recall, specificity, F1 score, Mean Squared Error (MSE), Mean Absolute Error (MAE), and R2. Results: RF-based SMIB achieved superior performance, with final imputation accuracy reaching up to 97% for categorical outcomes and strong numerical performance (R2 up to 0.999; MSE as low as 2.48 × 10−5). Sequential ordering and weighted bootstrapping improved stability in imbalanced clinical data under real-world missingness. Compared with the widely adopted Multiple Imputation by Chained Equations (MICE) approach, the proposed SMIB framework consistently demonstrated superior predictive performance across all evaluated categorical and numerical outcomes. Conclusions: The SMIB framework provides a robust and clinically aligned strategy for handling real-world missing values in MS datasets, improving imputation accuracy while preserving feature dependencies and feature relationships. The method supports reliable predictive analytics in healthcare contexts.
Missing binary predictors are common in reliability, quality control, and industrial decision systems, yet imputation methods are often chosen by convenience rather than evidence. We conduct a Monte Carlo study comparing mode substitution, sequential hot‐deck, missForest, MICE, and KNN with three neighbourhood sizes under MCAR, MAR, and MNAR missingness, across missingness rates from 5% to 50% and two predictor‐dependence structures. Performance is evaluated on three targets: exact recovery of missing binary cells, recovery of logistic‐regression coefficients, and downstream classification using logistic regression, naive Bayes, support vector machines, and random forests. The results reveal a clear trade‐off. KNN is strongest for exact cell recovery under MCAR and MAR, whereas missForest performs best under MNAR. MICE is the most reliable choice for downstream predictive performance across learners and missingness mechanisms. By contrast, mode imputation and sequential hot‐deck achieve the best coefficient recovery. The main implication is operational: in binary‐data environments, imputation should be chosen to match the analytical objective–reconstruction, inference, or prediction–because no single method dominates all targets simultaneously.
Manuel Delfino, Fabio Rapallo· Quality and Reliability Engi...· 0 citations
Early and accurate classification of melanoma is essential for improving patient outcomes and supporting clinical decision-making. Although numerous predictive models have been proposed, comparisons between classical statistical approaches and modern machine learning algorithms are often limited by heterogeneous analytical workflows and inconsistent validation strategies. This study aimed to compare the predictive performance of classical statistical and machine learning models for melanoma classification using a fully reproducible analytical framework. A retrospective observational study was conducted using the publicly available BCN20000 dermoscopic dataset from the ISIC Archive. After standardized data preprocessing, four routinely available clinical variables (age, sex, anatomical site and melanocytic status) were used to develop Logistic Regression, Generalized Additive Models, Random Forest and Extreme Gradient Boosting (XGBoost) classifiers. All models were trained and evaluated using the same stratified training/testing split, and their performance was assessed through discrimination, calibration and SHAP explainability analysis. Machine learning models, particularly XGBoost and Random Forest, achieved superior predictive performance compared with conventional statistical approaches, while patient age emerged as the most influential predictor of malignancy. The proposed framework provides a transparent and reproducible approach for objectively comparing predictive models and supports the development of accurate, interpretable, and reproducible clinical decision-support systems for melanoma classification.
Background: Breast cancer is the most common malignancy among women worldwide and a leading cause of cancer-related mortality. Despite advances in diagnosis and treatment, accurately predicting 5-year survival remains challenging because of disease heterogeneity. This study systematically evaluated the performance of machine learning (ML) models for predicting 5-year breast cancer survival and synthesised their overall discriminative performance. Methods: A systematic search of PubMed, Scopus, and Web of Science (2010–2024) identified studies developing ML models for 5-year survival prediction. Study quality was assessed using PROBAST. Logit-transformed area under the receiver operating characteristic curve (AUC) estimates were quantitatively synthesised using a random-effects meta-analysis with the restricted maximum likelihood (REML) estimator. Heterogeneity was assessed using Cochran’s Q and I2 statistics, publication bias using funnel plots and Egger’s regression test, and subgroup analyses according to study characteristics. Statistical analyses were performed using IBM SPSS version 29.0 and R version 4.6.1. Results: Fifteen studies were included. A wide range of ML models, including Random Forest, Random Survival Forest, and gradient boosting methods, was evaluated. The pooled analysis demonstrated good discriminative performance, with an overall AUC of 0.83 (95% CI: 0.80–0.86). Substantial heterogeneity was observed across studies. Ensemble-based models generally showed consistent performance. Publication bias was detected, and several studies exhibited moderate-to-high risk of bias. Conclusions: ML models show strong potential for predicting 5-year breast cancer survival and may support early risk stratification and clinical decision-making. However, greater methodological standardisation, rigorous external validation, and improved reporting are required before widespread clinical implementation.
A clear, statistically sound, yet easily understandable breast cancer diagnosis is a difficult issue in all healthcare systems, because early stages of breast cancer are critical in therapy success and long-term survivability. This machine-learning-based breast cancer classifier, in a statistically justified, rigorously experimentally validated way, classifies a set of 569 breast cancer cases with 9 cytological features for breast cancer diagnosis. The classifier uses a rigorous set of data cleanup measures, including missing-value substitution, correlation-based feature reduction, and projection into principal component space, to achieve high data quality, reduce redundancy, and enhance feature usefulness. Five supervised classifiers, in a widely accepted train-test model using an 80:20 random sample split and 5-fold cross-validation, are fitted and evaluated using Accuracy, Precision, Recall, F1-score, and Area under the ROC curve. In these tests, the Random Forest classifier got the best result, with 95.84% Accuracy, 95.31% Precision, 95.12% Recall, 95.21% F1-score and 0.982 area under the ROC curve; in a statistically sound consistency test using cross-validation, its mean accuracy reached 95.96% with a small standard deviation of 0.43. To provide a clear, interpretable indication of which features truly matter, we performed a feature-importance analysis on the best classifier, the Random Forest model. Results show that the expression levels of Bland Chromatin, Single Epithelial Cell Size, Normal Nucleoli, Uniformity of Cell Shape, Uniformity of Cell Size and Bare Nuclei are closely related to breast cancer diagnosis; this is almost the same as the clinical diagnosis findings, and very naturally suggests that abnormalities of cellular morphology and nuclei are major symptoms of breast cancer. In comparison, prior research may neglect validation and efficiency comparisons or focus only on the classifier's accuracy. Our method combines multiple levels of assessment (statistical data-by-data validation, feature importance, cross-validation, and comparison of different classifiers using ensemble learning) into a single evaluation system. This combined approach not only enhances predictive capability but also makes the entire setup more explicitly interpretable from a clinical perspective, thereby making it more suitable for health care decision support. Given the strong classification performance, interpretability, and validation suggested above, the model would help physicians detect breast cancer very early, reducing the risk of misdiagnosis.
T. Haripriya, M. V. Ramana Murthy, Ch. Vasavi et al.· International Journal of Eng...· 0 citations
Accurate breast cancer diagnosis is important for supporting early detection and improving clinical decision-making.. This study examines the theoretical and interpretative principles that support the machine learning algorithms applied to the Wisconsin Breast Cancer dataset and evaluates their ability to differentiate between benign and malignant neoplasms. To improve reliability, five-fold stratified cross-validation was additionally performed, and average performance values were considered. The logistic regression model achieved a high ROC-AUC value of 0.9960, whereas the performance of the random forest classifier was almost similar, with a slightly lower score. To enhance methodological transparency and interpretability, this study employs explainable AI techniques, including permutation importance, SHAP values, local and global explanation metrics, and partial dependence plots. These methods provide valuable insights into the extent to which tumour size and concavity contribute to malignancy prediction. A systematic study of computational complexity has been conducted, and the results show that the training and inference latency of logistic regression is significantly lower than that of random forests, indicating its potential suitability for future resource-constrained clinical decision-support investigations after external validation. The proposed framework achieved high discrimination performance while maintaining interpretability and low computational cost, suggesting its potential suitability for future clinical decision-support investigations after external validation.
Unknown authors· Journal of Computing & B...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
