GLOBAL VERSUS CLASS-BASED FEATURE RANKING FOR NOISY INSTANCE IDENTIFICATION IN BREAST CANCER DATA

Анотація

Abstract.

Background. The reliability of machine learning models in medical decision support systems depends not only on the classification algorithm but also on the quality of the training dataset. Previous studies evaluated the Coimbra Breast Cancer Dataset using a global consensus ranking of biomarkers obtained from multiple feature evaluators. However, such an approach assumes that biomarker relevance is identical for healthy individuals and breast cancer patients. Objective. This study investigates whether constructing independent biomarker rankings for each diagnostic class provides additional information about the dataset structure and improves the identification of potentially noisy patient records. Methods. The Coimbra Breast Cancer Dataset was divided into Healthy and Cancer subsets. For each subset, biomarker importance was independently evaluated using eight feature evaluation methods implemented in WEKA: Correlation Attribute Evaluation, Gain Ratio, Information Gain, OneR, ReliefF, SVM Attribute Evaluation, Symmetrical Uncertainty, and Classifier Attribute Evaluation. Consensus rankings were obtained by averaging the ranks assigned by the individual evaluators. The resulting class-specific rankings were compared with the conventional global ranking. Potentially noisy records were further analyzed using the Interquartile Range filter. Results. Statistical analysis confirmed that several biomarkers exhibit different distribution characteristics in the two diagnostic classes. Although the Healthy and Cancer subsets produced similar consensus rankings, both differed substantially from the global ranking. In particular, MCP-1 changed from the lowest-ranked biomarker in the global analysis to the highest-ranked biomarker in both class-specific rankings, whereas Age became the least informative feature after class separation. The proposed approach identified five potentially noisy records in each diagnostic class. Most of these records were not statistical outliers according to the interquartile range criterion, indicating that they represent atypical combinations of biomarker values rather than isolated extreme measurements. Conclusions. The obtained results demonstrate that class-specific feature ranking provides additional information about the internal structure of medical datasets and complements conventional global feature ranking. The proposed approach can be used as an exploratory preprocessing technique for identifying potentially atypical patient records before applying machine learning methods.

References

1) Topol, E. J. (2019). High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine, 25(1), 44–56. https://doi.org/10.1038/s41591-018-0300-7

2) Austria, Y. D., Goh, M. L., Sta. Maria Jr., L., Lalata, J.-A., Goh, J. E., & Vicente, H. (2019). Comparison of Machine Learning Algorithms in Breast Cancer Prediction Using the Coimbra Dataset. International journal of simulation: systems, science & technology. https://doi.org/10.5013/ijssst.a.20.s2.23

3) Yue, J., Zhao, N., & Liu, L. (2020). Prediction and Monitoring Method for Breast Cancer: A Case Study for Data from the University Hospital Centre of Coimbra. Cancer Management and Research, Volume 12, 1887–1893. https://doi.org/10.2147/cmar.s242027

4) Alfian, G., Syafrudin, M., Fahrurrozi, I., Fitriyani, N. L., Atmaji, F. T. D., Widodo, T., Bahiyah, N., Benes, F., & Rhee, J. (2022). Predicting Breast Cancer from Risk Factors Using SVM and Extra-Trees-Based Feature Selection Method. Computers, 11(9), 136. https://doi.org/10.3390/computers11090136

5) Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, M., Chou, K., Cui, C., Corrado, G., Thrun, S., & Dean, J. (2019). A guide to deep learning in healthcare. Nature Medicine, 25(1), 24–29. https://doi.org/10.1038/s41591-018-0316-z

6) Beam, A. L., & Kohane, I. S. (2018). Big Data and Machine Learning in Health Care. JAMA, 319(13), 1317. https://doi.org/10.1001/jama.2017.18391

7) García, S., Luengo, J., & Herrera, F. (2015). Data Preprocessing in Data Mining. Springer International Publishing. https://doi.org/10.1007/978-3-319-10247-4

8) Pei, J., Han, J., & Kamber, M. (2011). Data Mining: Concepts and Techniques. Elsevier Science & Technology Books.

9) Hodge, V., & Austin, J. (2004). A Survey of Outlier Detection Methodologies. Artificial Intelligence Review, 22(2), 85–126. https://doi.org/10.1023/b:aire.0000045502.10941.a9

10) Zhu, X., & Wu, X. (2004). Class Noise vs. Attribute Noise: A Quantitative Study. Artificial Intelligence Review, 22(3), 177–210. https://doi.org/10.1007/s10462-004-0751-8

11) Frenay, B., & Verleysen, M. (2014). Classification in the Presence of Label Noise: A Survey. IEEE Transactions on Neural Networks and Learning Systems, 25(5), 845–869. https://doi.org/10.1109/tnnls.2013.2292894

12) Chuiko, G., & Honcharov, D. (2025). Breast Cancer Dataset from Coimbra: Pre‑Ratings of Its Value to Machine Learning and Diagnosis. Digital Technologies Research and Applications, 4(2), 182–193. https://doi.org/10.54963/dtra.v4i2.1348

13) Saeys, Y., Inza, I., & Larranaga, P. (2007). A review of feature selection techniques in bioinformatics. Bioinformatics, 23(19), 2507–2517. https://doi.org/10.1093/bioinformatics/btm344

14) Patrício, M., Pereira, J., Crisóstomo, J., Matafome, P., Gomes, M., Seiça, R., & Caramelo, F. (2018). Using Resistin, glucose, age and BMI to predict the presence of breast cancer. BMC Cancer, 18(1). https://doi.org/10.1186/s12885-017-3877-1

15) Patrcio, M.; Pereira, J.; Crisstomo, J.; et al., 2018. Breast Cancer Coimbra. UCI Machine Learning Repository. Available from: https://archive.ics.uci.edu/dataset/451/breast+cancer+coimbra