Analyzing Ensemble Bias and Error Distributions for Robust Stroke Prediction 1

Read full text Summarise with VIRAT
Authors: Soheli, Jahan, Student, Abdul Alim, Prianka Mondal, Reduanur Rahman, Hena Joti, Mobarok Hossen, Bappi Bhushan Das, Al-Amin, Md, Hadi, Ahmed Faizul
Cited by: 0
References: 12

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Stroke is among the most prevalent causes of disability and death worldwide and is also a significant health burden in Bangladesh. Timely and accurate predictions of stroke in practice is essential, but there are constraints to conventional evaluations in terms of data and interpretation. In this project, we develop and evaluate supervised machine learning models and newer ensemble methods for predicting strokes based on a newly curated benchmark database containing 5110 patient records with 12 demographics, clinical, and lifestyle features. Preprocessing was conducted in line with modern healthcare analytics data handling procedures like imputation, one-hot encoding, and stratified cross-validation. In addition to singular classifiers, hybrid and cross-paradigm ensembles were employed like Naïve Bayes + CatBoost, KNN + Random Forest, TabNet + LightGBM) to utilize probabilistic, boosting, and deep learning tabular models. Five performance metrics were used to assess the final model: macro precision and recall, rigorous accuracy values, ROC-AUC, and balanced F1-scores. Boosting with Naïve Bayes (AdaBoost) showed a strong balanced classifier (precision 0.7861 , recall 0.5189, accuracy 0.9041) and when we blended GBM + RobustBoost we achieved the highest average recall score (0.7312), while the LGBM + TabNet ensemble classifier delivered the best overall accuracy (0.8992), with a competitive ROC-AUC (0.790). From our confusion matrix analysis we found boosting primarily minimized false positives while leaving a greater cohort of missed positives, bagging achieved perfect recall at the expense of high total false positives, which inflated accuracy.

Fields of Study & Subjects

Artificial Intelligence in Healthcare Imbalanced Data Classification Techniques Data Mining and Machine Learning Applications

Publication details

Journal / source
HAL
Publication date
26 February 2026
Language
EN