Disease Prediction Using Machine Learning

by: Tomás Ospina.

Overview

A supervised-learning study that predicts a disease from a set of reported symptoms, combining three classifiers into a majority vote rather than trusting any one of them.

Pipeline: dataset, oversampling, cross-validated SVM / Naive Bayes / Random Forest, majority vote

Dataset & Preprocessing

2,000 records with ten binary symptom features — fever, headache, nausea, vomiting, fatigue, joint pain, skin rash, cough, weight loss and yellow eyes — and the diagnosed disease as the target.

Two preprocessing steps do most of the work: a LabelEncoder on the target, and a RandomOverSampler to correct the class imbalance, since the rarer diseases would otherwise be predicted away entirely by a model optimising raw accuracy.

Modelling

Three classifiers are trained and validated independently with StratifiedKFold cross-validation — keeping the class distribution intact in every fold — and their predictions are combined by majority vote:

Model	Role in the ensemble
SVM (`SVC`)	Margin-based separation of the symptom space
Gaussian Naive Bayes	Fast probabilistic baseline over the features
Random Forest	Non-linear feature interactions, robust to noise

Results are read off a confusion matrix per model plus the ensemble, so a disease that one classifier consistently confuses is visible rather than buried in an accuracy number.

What I Learned

The oversampling step changed the results more than the model choice did. On an imbalanced medical dataset, a high accuracy score usually means the model has learned to ignore the rare classes — which is exactly the failure mode that matters here.

🔒 Private repository — happy to walk through the code on request.