Scalable Identification of Clinically Relevant Chronic Obstructive Pulmonary Disease Documents in Large-Scale...

Area of Science:

• Clinical Informatics
• Machine Learning
• Natural Language Processing

Background:

• Electronic health records generate vast amounts of clinical notes, posing challenges for data analysis.
• Learning algorithms and large language models are sensitive to noise and irrelevant data, leading to performance degradation and inaccuracies.
• Efficiently filtering millions of clinical documents for relevance is crucial, especially in resource-constrained environments.

Purpose of the Study:

• To develop a novel framework for determining document relevance in clinical settings.
• To address the challenge of efficiently filtering large volumes of clinical notes for advanced language model processing.
• To improve the accuracy and efficiency of information retrieval from electronic health records.

Main Methods:

• Developed a framework using weak supervision and domain-expert heuristics to create "silver standard" labels.
• Utilized various text representation techniques including bag of words, TF-IDF, lightweight document embeddings, and UMLS concept extraction.
• Trained and evaluated random forest, extreme gradient boosting, and k-nearest neighbor classifiers on expert-annotated and held-out test datasets.

Main Results:

• The combination of lightweight document embedding and a random forest classifier achieved the highest performance.
• This model demonstrated a precision of 0.73, recall of 0.86, and F1-score of 0.80 for identifying relevant chronic obstructive pulmonary disease (COPD) documents.
• The proposed framework significantly outperformed baseline heuristics and other tested methods.

Conclusions:

• A novel framework for identifying COPD-relevant clinical documents was presented, utilizing lightweight embedding and machine learning.
• This approach effectively filters pertinent documents, enhancing information retrieval precision and scalability.
• The framework's minimal annotation needs show promise for diverse healthcare applications, potentially optimizing clinical outcomes through efficient document selection.