Medical subdomain classification of clinical notes using a machine learning-based natural language processing...

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Classifying clinical note medical subdomains is crucial for machine learning applications.
• Developing accurate medical subdomain classifiers requires robust natural language processing (NLP) pipelines.

Purpose of the Study:

• To construct and evaluate a machine learning-based NLP pipeline for accurate medical subdomain classification.
• To develop and compare medical subdomain classifiers using various data representation and learning algorithms.

Main Methods:

• Utilized the clinical Text Analysis and Knowledge Extraction System (cTAKES) and Unified Medical Language System (UMLS).
• Extracted features from two diverse clinical note datasets (iDASH and MGH).
• Built and evaluated classifiers using different data representation methods and supervised learning algorithms, including deep learning and shallow learning models.

Main Results:

• A convolutional recurrent neural network achieved the highest performance (AUC 0.975-0.991, F1 0.845-0.870).
• A linear support vector machine with hybrid features (BoW and UMLS concepts) offered strong performance (AUC 0.957-0.964, F1 0.932-0.934) with better clinical interpretability.
• Classifiers demonstrated portability across datasets, achieving an F1 score threshold of 0.7 for half of the subdomains.

Conclusions:

• Supervised learning-based NLP approaches are effective for developing medical subdomain classifiers.
• Deep learning models offer superior performance, while shallow learning models provide comparable results with enhanced clinical interpretability.
• Developed classifiers exhibit portability across datasets from different institutions.