Automated Multilabel Mpox Research Classification with Explainable Transformer Models

As Mpox research continues to expand rapidly, organizing this information for policymakers and healthcare workers has become increasingly difficult. This study aims to automate the categorization of Mpox research articles using multilabel classification. By applying transformer-based models, the author seeks to classify abstracts into three primary domains: outbreak-related, prevention-related, and vaccine-related research.

The researcher collected 14,590 research abstracts from PubMed using a set of nine Mpox-related keywords. After preprocessing the text—including stop-word removal, lemmatization, and stemming—the data was used to fine-tune several transformer architectures, including BERT, RoBERTa, ALBERT, DistilBERT, and ELECTRA. The study employed a multilabel approach, acknowledging that a single research paper often addresses multiple topics simultaneously. To ensure transparency, the author integrated SHAP (SHapley Additive exPlanations) to visualize which specific words and phrases most heavily influenced the model's classification output.

Among the tested models, BERT demonstrated the highest performance, achieving an accuracy of 97.05%, a micro F1 score of 97.67%, and a macro F1 score of 96.46%. The results indicate that transformer models are highly effective at handling overlapping research topics in the biomedical domain. The application of SHAP successfully provided insights into the model's decision-making process, confirming that the classification was driven by relevant medical terminology rather than noise.

Automating the classification of scientific literature is essential for public health preparedness. By providing a reliable, explainable tool to organize Mpox research, this work helps researchers and health authorities quickly synthesize information, identify gaps in current knowledge, and accelerate the development of interventions during public health emergencies.