Scalable Identification of Clinically Relevant COPD Documents: A Lightweight NLP Model for Large-Scale EHR Datasets

The widespread adoption of electronic health records (EHRs) has resulted in the generation of large volumes of clinical notes. Learning algorithms and large language models (LLMs) train from these resources but are susceptible to noise—irrelevant or non-informative data from them. This sensitivity can lead to significant challenges, including performance degradation and the generation of inaccurate predictions or “hallucinations.” This study addresses a critical challenge in clinical informatics: efficiently filtering millions of documents for relevance before advanced language model processing, particularly in resource-constrained environments. We present a novel framework for determining document relevance in clinical settings, utilizing a chronic obstructive pulmonary disease (COPD) dataset. We developed a novel framework using weak supervision and domain-expert heuristics to generate “silver standard” labels for training data and expert annotated labels (gold stand),creating two datasets to optimize the model during the development phase and subsequent testing phase. Various text representation techniques, including Bag-of-Words, TF-IDF, lightweight document embeddings, compression-based features, and UMLS concept extraction, were evaluated. These representations were used to train Random Forest, XGBoost, and K-Nearest Neighbors classifiers. Models were optimized on a small expert-annotated dataset and evaluated on a held-out test set. The combination of lightweight document embedding with a Random Forest classifier demonstrated the best performance, achieving a precision of 0.75, recall of 0.89, and F1-score of 0.81 (95% CI: 0.76-0.87) for identifying relevant COPD documents. This significantly outperformed baseline heuristics (precision: 0.70, recall: 0.38, F1-score: 0.50, 95% CI: 0.43-0.56) and other tested methods. Our study presents a novel framework for identifying COPD-relevant clinical documents using lightweight embedding and machine learning. This approach effectively filters pertinent documents, enhancing information retrieval precision. The framework’s scalability and minimal annotation needs make it promising for diverse healthcare applications, potentially optimizing clinical outcomes through efficient document selection for data-driven decision support systems.