Survival prediction
with the classics.

Abstract

Binary survival classification on the MIMIC clinical dataset using KNN and SVM. The deliberate choice of classical over deep methods is the point: the features are tabular, the dataset is small, and interpretability matters.

Method

Feature engineering over vitals and labs → KNN & SVM with grid-searched hyperparameters → class-imbalance handling → ROC / AUC evaluation.

Why not deep?

Clinical tabular data, in small samples, rewards models with strong inductive bias. A well-tuned SVM with an RBF kernel outperformed every neural baseline I tried and — crucially — offered coefficients a clinician could actually read.

Takeaway

Sometimes the right answer is the boring one. Understanding when deep methods are the wrong tool is as important as knowing how to fine-tune them.