When the data come as structured tables rather than images, classic algorithms often outperform more exotic ones — and they are easier to verify and defend in a clinic. Five of them are the standard workhorses behind diagnostic decision support.
Logistic regression predicts the probability that a case belongs to a class, and its coefficients are readable: a doctor can see which features push the risk up. Decision trees encode decisions as branches that anyone can follow, though deep trees tend to overfit.
Support vector machines draw the decision boundary — the hyperplane — that best separates classes, and use the kernel trick for complex, non-linear patterns. K-nearest neighbours classifies by the majority vote of the closest examples: simple and effective as a baseline. Naive Bayes applies the calculus of probabilities efficiently, especially on small datasets.
How they behave on breast-tumour data
- A public dataset of digitised cell samples separates benign from malignant very cleanly
- Simple models such as logistic regression can reach near-perfect discrimination
- Choosing an algorithm is a matter of data, interpretability and robustness