Skip to content

From stroke prediction to three-tier risk stratification

Status
Completed
Date
2026
Venue
MS AI coursework, Bahria University
Tags
Healthcare ML · Imbalanced learning · Model comparison

  • 92.37%

    Random Forest accuracy

    vs. 86.89% for the ANN

  • 0.71

    macro F1

    0.98 Low, 0.90 Moderate

  • 511

    unseen test records

Abstract

A yes/no stroke label is too late to act on. This project re-engineers the Kaggle stroke dataset into a three-tier risk scale for earlier intervention, with a leakage-safe preprocessing pipeline, and compares a Random Forest ensemble with a multi-layer perceptron on 511 unseen test records.

01Pipeline

  • Median imputation for missing biomedical values, one-hot encoding that avoids the dummy-variable trap, and z-score scaling.
  • Multiclass SMOTE applied only to the stratified training split (7,194 balanced records), so the 1,022 tuning and 511 test rows stay untouched.
  • Models: a 150-tree Random Forest (Gini) and an MLP with ReLU and softmax.

02What it says

On tabular clinical data the ensemble beats the neural network clearly. The High-risk tier is still the hardest to get right, which is why macro F1 sits at 0.71.