Patient-Aware Automl Radiomics for Multiclass Pneumonia Classification from Chest X-Ray Images
Journal of Al-Farabi for Engineering Sciences · 01.12.2025 · 4 (2) · 471-489
DOI: 10.59746/kh2a0q71 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Al-Farabi for Engineering Sciences, online erschienen am 01.12.2025. Verfasst von Israa Faisal Jassam.
Abstract (Originalsprache)
Pneumonia remains a major respiratory disease, and accurate differentiation among bacterial pneumonia, viral pneumonia, and normal chest radiographs is essential for timely clinical decision-making. This study proposes a patient-aware AutoML radiomics framework for multiclass pneumonia classification using precomputed quantitative features extracted from chest X-ray images. The dataset comprised 3,527 samples, including 1,469 bacterial pneumonia, 1,218 normal, and 840 viral pneumonia cases, representing 2,431 unique patients. Patient identifiers were extracted from image filenames and used to perform a 70/30 group-based split, producing 2,477 training samples from 1,701 patients and 1,050 independent test samples from 730 patients, with no patient overlap between the two subsets. After removing metadata, non-numeric, zero-variance, redundant, and uninformative variables, 1,107 radiomic features were retained. To address class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set, while the test set remained unchanged. Seven AutoGluon classifiers, including neural-network, Random Forest, Extra Trees, and LightGBM models, were trained and compared using macro-averaged F1-score. PyTorch Neural Network achieved the best performance, with 99.62% accuracy, 99.67% macro-precision, 99.56% macro-recall, and 99.61% macro-F1. FastAI Neural Network achieved a comparable macro-F1 of 99.47%, whereas the best classical model, Random Forest Gini, achieved 90.73%. By preserving the original test distribution, the evaluation provides a more realistic estimate of generalisation to previously unseen patients and three clinically imbalanced diagnostic categories. The findings indicate that neural tabular models can effectively capture complex nonlinear relationships within high-dimensional radiomic data. The proposed framework provides a leakage-aware and reproducible approach for multiclass pneumonia diagnosis; however, external and multi-centre validation is still required before clinical deployment.
Zitieren
APA
Jassam, I. F. (2025). Patient-Aware Automl Radiomics for Multiclass Pneumonia Classification from Chest X-Ray Images. Journal of Al-Farabi for Engineering Sciences, 4(2), 471-489. https://doi.org/10.59746/kh2a0q71
Vancouver
Jassam IF. Patient-Aware Automl Radiomics for Multiclass Pneumonia Classification from Chest X-Ray Images. Journal of Al-Farabi for Engineering Sciences. 2025;4(2):471-489. doi:10.59746/kh2a0q71
BibTeX
@article{Jassam2025,
author = {Jassam, Israa Faisal},
title = {Patient-Aware Automl Radiomics for Multiclass Pneumonia Classification from Chest X-Ray Images},
journal = {Journal of Al-Farabi for Engineering Sciences},
year = {2025},
volume = {4},
number = {2},
pages = {471--489},
doi = {10.59746/kh2a0q71}
}
EUVIMED