Advancing Clinical Diagnosis Through Comparative Analysis of Machine Learning and Transformer-Based Models for Hepatitis B Detection

Abstract

Early and accurate diagnosis of hepatitis B remains a critical issue in low-resource healthcare settings, where reliance on straightforward laboratory tests might limit the sensitivity of diagnosis. This research explores the prediction of hepatitis B infection based on conventional biochemical markers and demographic data by four machine learning models: XGBoost, LightGBM, TabPFN, and TabKANet. The goals are twofold: first, to compare which model type is best, the Gradient Boosted Decision Trees (GBDTs) or transformer-based models, and second, to compare the feasibility of such models being included in clinical workflows. The models were evaluated against a real-world dataset of a Ghanaian hospital-based screening program. Performance was calculated in terms of accuracy, precision, recall, F1-score, and area under the curve (AUC). Results showed that transformer models, particularly TabKANet, outperformed traditional GBDT models in recall (97.0%), F1-score (65.5%), and AUC (0.92), with high ability to detect true positive cases. Confusion matrix analyses also confirmed the validity of TabKANet in minimizing false negatives, which is very much required in clinical diagnosis. These findings support the application of artificial intelligence-based technology in clinical laboratories as a second confirmatory system for disease diagnosis. The study also encourages future research into the use of transformer models in healthcare, especially in settings where confirmatory testing is unavailable. Received: 7 August 2025 | Revised: 23 December 2025 | Accepted: 20 January 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The dataset used in this study is publicly available and was accessed in compliance with its original ethical approvals and usage terms. The data can be accessed via the bioRxiv preprint server at https://doi.org/10.1101/2024.05.01.24306678. Access and use of the dataset were in accordance with bioRxiv's policies for public preprint content, ensuring ethical and permitted usage for research purposes. Author Contribution Statement Akinyemi Omololu Akinrotimi: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision, Project administration. Israel Oluwabusayo Omotosho: Resources, Data curation. Olugbenga Olayinka Owolabi: Resources, Data curation. Oluwaseun Adewale Olubunmi: Validation, Resources, Data curation. Ibrahim Garba: Validation, Formal analysis, Investigation, Writing – review & editing. Ndie Ngalame Dionysius: Validation, Formal analysis, Investigation, Writing – review & editing. Early and accurate diagnosis of hepatitis B remains a critical issue in low-resource healthcare settings, where reliance on straightforward laboratory tests might limit the sensitivity of diagnosis. This research explores the prediction of hepatitis B infection based on conventional biochemical markers and demographic data by four machine learning models: XGBoost, LightGBM, TabPFN, and TabKANet. The goals are twofold: first, to compare which model type is best, the Gradient Boosted Decision Trees (GBDTs) or transformer-based models, and second, to compare the feasibility of such models being included in clinical workflows. The models were evaluated against a real-world dataset of a Ghanaian hospital-based screening program. Performance was calculated in terms of accuracy, precision, recall, F1-score, and area under the curve (AUC). Results showed that transformer models, particularly TabKANet, outperformed traditional GBDT models in recall (97.0%), F1-score (65.5%), and AUC (0.92), with high ability to detect true positive cases. Confusion matrix analyses also confirmed the validity of TabKANet in minimizing false negatives, which is very much required in clinical diagnosis. These findings support the application of artificial intelligence-based technology in clinical laboratories as a second confirmatory system for disease diagnosis. The study also encourages future research into the use of transformer models in healthcare, especially in settings where confirmatory testing is unavailable. Received: 7 August 2025 | Revised: 23 December 2025 | Accepted: 20 January 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The dataset used in this study is publicly available and was accessed in compliance with its original ethical approvals and usage terms. The data can be accessed via the bioRxiv preprint server at https://doi.org/10.1101/2024.05.01.24306678. Access and use of the dataset were in accordance with bioRxiv's policies for public preprint content, ensuring ethical and permitted usage for research purposes. Author Contribution Statement Akinyemi Omololu Akinrotimi: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision, Project administration. Israel Oluwabusayo Omotosho: Resources, Data curation. Olugbenga Olayinka Owolabi: Resources, Data curation. Oluwaseun Adewale Olubunmi: Validation, Resources, Data curation. Ibrahim Garba: Validation, Formal analysis, Investigation, Writing – review & editing. Ndie Ngalame Dionysius: Validation, Formal analysis, Investigation, Writing – review & editing.