Text mining in biomedical literature: identifying chemical and disease relationships

Abstract

Text mining is an essential method for extracting valuable information from biomedical literature, especially concerning chemical-disease relationships (CDR). This study integrates named entity recognition (NER), relation extraction (RE), classification, and topic modeling to analyze kratom-related research articles from PubMed. Using the biocreative V chemical-disease relation (BC5CDR) dataset, NER achieved high recall and accuracy, demonstrating its ability to identify diverse entities. However, low precision revealed challenges with false positives due to ambiguous contexts and term overgeneralization. Network analysis of extracted relationships identified “Neuropathic” as the most central entity, underscoring its importance in linking chemicals like “mitragynine” to clinical conditions. Classification with bidirectional encoder representations from transformers for biomedical text mining (BioBERT) yielded moderate results, with an F1-score of 0.48, precision of 0.43, and recall of 0.38, highlighting the need for larger, curated datasets to improve accuracy. Bidirectional encoder representations topic modeling (BERTopic) uncovered five main topics with moderate coherence (C_V = 0.5567), emphasizing themes related to kratom’s toxicity and clinical implications, although capturing semantic relationships remained challenging. This research demonstrates the potential of text mining in biomedical knowledge discovery, revealing key insights and thematic patterns in scientific literature. Improved data preprocessing, model fine-tuning, and manual validation are recommended to enhance the reliability and application of these methods in future studies.