Natural Language Processing in Electronic Health Record Mining for Clinical Decision Support

Abstract

An investigation concerning an NLP-based approach applied to EHR mining with improved clinic decision support. Using TF-IDF, WordEmbeddings, NER, and LDA, the study seeks to leverage clinical narratives for critical analysis on the topic of unplanned readmission using administrative data only. A diverse dataset consisting of de-identified EHR is standardized and goes through proper preprocessing for use as input for each algorithm. TF-IdF works well for term extraction; Word2Vec reflects semantic relationships; NER is precise about medical entities; and LDA suggests that there are hidden thematic patterns among the variables in the data. This comparison as well as alignment with other works reveals the sophisticated superiority of every algorithm from different data and linguistic cases. These results illustrate the strength of NLP for revealing important data, which contributes to a wider discussion about its use in health care practices. Interestingly, TF-IDF reaches 0.85 precision and 0.92 recall for the core-medical terms detection. The semantic understanding of word2vec is revealed in a higher cosine similarity which is 0.78 for these phrases – "Diabetes" vs. "Insulin". In medical condition identification, NER obtains 0.92, 0.88, and 0.90 for precision, recall, and F1-score. Coherence scores of 0.75 and 0.82 indicate that LDA indeed discovers underlying thematic structures for different topics.