Class-Imbalance Aware Learning for Enhanced Noninvasive Neonatal Jaundice Diagnosis – Using Convnext-Tiny with Focal Loss (CNXT-FL)
DOI:
https://doi.org/10.23055/ijietap.2026.33.4.11505Keywords:
Neonatal Jaundice Diagnosis, Transformer-Inspired ConvNeXt-Tiny, , Focal Loss, Class Imbalance Handling, AI-Based Image Analysis, Non-Invasive Medical Diagnosis, Deep Learning in HealthcareAbstract
Early diagnosis of neonatal jaundice carries significance as it could prevent serious neurological problems. Conventional diagnostics based on assessment of total serum bilirubin (TSB) in blood samples are invasive, painful, time-consuming, and require a laboratory setting. As such facilities are often obscure in resource-limited remote locations, AI-based image analysis was frequently investigated as an alternate solution to optimize clinical workflow. However, the need for an optimized healthcare decision support system remains a research gap. This research introduces a Transformer-inspired ConvNeXt-Tiny architecture with Focal Loss (CNXt-Fl) for rapid, scalable, and accurate diagnosis of neonatal jaundice. The proposed framework includes (i) Albumentations-based augmentation for ensuring model robustness to image acquisition conditions, (ii) ‘focal loss’ for class-weighted learning to address class imbalance, and (iii) a pretrained ConvNeXt-Tiny model to ensure insensitivity to small datasets. Experimental results from CNXt-Fl evaluated on a public neonatal jaundice dataset consisting of 760 neonatal images achieved an accuracy and F1-score of 0.9912 and 0.99, respectively. Further, comparison with conventional CNN, other pre-trained models, and state-of-the-art deep learning models revealed that the proposed model achieves better results comparatively. These results highlight the potential of lightweight transformer-inspired architectures with class-imbalance aware learning for reliable early detection of neonatal jaundice in real-world healthcare environments. Thus, this work holds significance in the optimization of healthcare decision support systems.
Published
How to Cite
Issue
Section
License
The Author(s) must formally transfer each article's copyright before publication in the INTERNATIONAL JOURNAL OF INDUSTRIAL ENGINEERING. Such transfer enables the Journal to defend itself against plagiarism and other forms of copyright infringement. Your cooperation is appreciated.
You agree that the copyright of your article to be published in the INTERNATIONAL JOURNAL OF INDUSTRIAL ENGINEERING - THEORY, APPLICATIONS, AND PRACTICE is hereby transferred, throughout the World and for the full term and all extensions and renewals thereof, to INTERNATIONAL JOURNAL OF INDUSTRIAL ENGINEERING - THEORY, APPLICATIONS, AND PRACTICE.
The Author(s) reserve(s): (a) the trademark rights and patent rights, if any, and (b) the right to use all or part of the information contained in this article in future, non-commercial works of the Author's own, or, if the article is a "work-for-hire" and made within the scope of the Author's employment, the employer may use all or part of the information contained in this article for intra-company use, provided the usual acknowledgments are given regarding copyright notice and reference to the original publication.
The Author(s) warrant(s) that the article is Author's original work and has not been published before. If excerpts from copyrighted works are included, the Author will obtain written permission from the copyright owners and credit the article's sources.
The author also warrants that the article contains no libelous or unlawful statements and does not infringe on the rights of others. If the article was prepared jointly with other Author(s), the Author agrees to inform the co-Author(s) of the terms of the copyright transfer and to sign on their behalf; or in the case of a "work-for-hire," the employer or an authorized representative of the employer.
The journal does not provide the author copy of the final paper when it is published. The author(s) can make(s) a subscription to INTERNATIONAL JOURNAL OF INDUSTRIAL ENGINEERING - THEORY, APPLICATIONS, AND PRACTICE if they want to get the final paper that has already been published.
The journal is registered with the Library of Congress (ISSN # 1943-670X). All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the journal.
The author reserves patent and trademark rights and the right to use all or part of the information contained in the article in future non-commercial works.
