What question did this study set out to answer?

The aim is to analyze COPD severity classification using AI-driven algorithms compared to traditional methods.

April 1, 2026International Journal Of Informative and Futuristic Research0 citationsOpen Access

COPD Datasets Analysis Using AI-Driven Algorithms

SS.Parthasarathi VNV.Lakshmi Narasimhan DD.Rammohan

Key Points

The aim is to analyze COPD severity classification using AI-driven algorithms compared to traditional methods.
Employ comparative analysis of three gradient boosting algorithms: XGBoost, LightGBM, and CatBoost.
Benchmark against a baseline model using a 1D Convolutional Neural Network.
Implement a Stacking Ensemble approach combining all gradient boosting models.
Conduct experiments on the Kaggle COPD Student Dataset with 101 patients and 19 clinical features.
Perform five-fold cross-validation for model validation.
XGBoost achieved the highest accuracy of 90.48% and an AUC-ROC of 96.83%.
CatBoost exhibited the best discriminative ability with a near-perfect AUC-ROC of 99.31%.
LightGBM provided stable results with 85.71% accuracy and 95.08% AUC-ROC.
The CNN baseline significantly underperformed with only 42.86% accuracy.
SHAP analysis identified FEV1PRED, FEV1, and FVCPRED as key clinical predictors.

Abstract

Chronic Obstructive Pulmonary Disease (COPD) is a progressive respiratory condition affecting over 300 million people globally and remains significantly underdiagnosed due to reliance on traditional spirometry-based screening. This paper presents a comparative study of three state-of-the-art gradient boosting algorithms-XGBoost, LightGBM, and CatBoost-for automated COPD severity classification on real clinical tabular data, benchmarked against a 1D Convolutional Neural Network (CNN) baseline. Additionally, a Stacking Ensemble combining all three gradient boosting models via a Logistic Regression meta-learner is proposed as a novel hybrid approach. Experiments are conducted on the Kaggle COPD Student Dataset comprising 101 patients and 19 clinical features. The experimental evaluation demonstrates that gradient boosting algorithms significantly outperform the 1D CNN baselines for COPD severity classification using small clinical tabular data. Among all models, XGBoost achieved the best overall performance with 90.48% accuracy, 89.76% weighted F1-score, and 96.83% AUC-ROC, indicating superior predictive capability and balanced classification across severity levels. CatBoost achieved the highest discriminative ability with a near-perfect 99.31% AUC-ROC, while LightGBM delivered stable and competitive results with 85.71% accuracy and 95.08% AUC-ROC. In contrast, the CNN baseline performed poorly, achieving only 42.86% accuracy and 25.71% F1-score, confirming that deep learning architectures are less effective for small structured datasets. The Stacking Ensemble achieves competitive performance (85.71% accuracy, 97.9% AUC-ROC), confirming ensemble combination as a viable research direction. Five-fold cross-validation further confirmed model robustness, with gradient boosting models achieving mean F1-scores above 83%, substantially higher than the CNN baseline. SHAP explainability analysis consistently identified FEV1PRED, FEV1, and FVCPRED as the most influential clinical predictors, aligning with established COPD severity guidelines and supporting the clinical validity of the models.

Perguntar à IA

Bookmark

View Full Paper