A clinical named entity recognition model using pretrained word embedding and deep neural networks

Clinical Named Entity Recognition (NER) within Electronic Medical Records (EMRs) has seen substantial research attention. Since much clinical information resides in unstructured text, NER technology is pivotal in extracting vital patient data from such sources. The ubiquity of EMRs has fueled interest in leveraging technology for natural language processing, particularly in the domain of Biomedical Named Entity Recognition (Bio-NER), which is challenged by a range of entities like genes, proteins, medications, and diseases. Recent Natural Language Processing (NLP) advancements have demonstrated remarkable performance through text encoder pre-training. A linchpin in the efficacy of neural sequence labeling models is the selection and encoding of input data, which is crucial for generating nuanced semantic and morphological word vectors. This paper emphasizes one of the variants of the recurrent neural network model and Conditional random field for improving the performance of the proposed model. This study addresses the formidable task of Bio-NER through a deep neural network model by incorporating biological corpus statistics and refined hyperparameters. The embedding layer converts numerical sequences into dense word embeddings by passing sequences through an embedding matrix comprising parameter weights. Pre-trained word embeddings like word2vec or GloVe facilitate embedding creation based on input. The proposed model achieves an impressive F1-Score of 88.34%, surpassing the need for post-processing or heuristic rules. Instead, it harnesses pre-trained word embeddings to streamline the training process. Besides reducing training complexities, it also enhances the accuracy and effectiveness of the Bio-NER and detection task.