Predicting breast cancer recurrence with machine learning classifiers
- Article
- Health
In 2020 the World Health Organization counted nearly 2.3 million women diagnosed with breast cancer, making it one of the most common cancers among women and the leading cause of cancer death for them. Early detection is the most effective tool against it, and that applies to recurrence as well as to first diagnosis. Vincent Peter Magboo and Ma. Sheila Magboo, in a 2021 paper in Procedia Computer Science, tested how well standard machine learning classifiers can predict whether a patient’s cancer will come back.
The data
They used the Wisconsin Prognostic Breast Cancer (WPBC) dataset, a public collection of records that includes whether each case recurred along with features measured from digitised images of breast mass samples: the dimensions of cell nuclei, tumour size, and lymph node status among them. It is a small dataset, which makes the choice of method and preprocessing matter more than it would with millions of rows.
The experiment
Four classifiers were compared: logistic regression, Naive Bayes, k-nearest neighbours, and support vector machines. Each was trained with 10-fold cross validation under four data configurations:
- scaling only,
- scaling with principal component analysis (PCA),
- scaling with PCA and random oversampling of the training set, to balance the recurrence and non-recurrence classes, and
- removing features with more than 80 percent correlation and keeping the rest.
Four algorithms by four configurations gave 16 model fits, and the authors compared them to find the best combination.
Findings
Logistic regression performed best in most configurations, followed by support vector machines and k-nearest neighbours. Naive Bayes did poorly by comparison.
Configuration mattered, and not uniformly. Support vector machine performance stayed roughly constant across all four setups, while k-nearest neighbours improved slightly under scaling with PCA and oversampling. Preprocessing steps that help one algorithm may do nothing for another, which is a reason to test combinations rather than settle on a pipeline first and an algorithm second.
What the authors take from it
Magboo and Magboo suggest that a logistic regression model of this kind could help physicians detect likely recurrence and propose treatment on the basis of the predicted prognosis. The post framed it as an example of data mining and model building applied to a problem where early detection is the whole game. It is also a reminder that the simplest model in the comparison, a logistic regression, was the one that worked best.