Word2Vec inversion and traditional text classifiers for phenotyping lupus.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Machine Learning in Healthcare

Background:

• Manual chart review of electronic health records (EHR) is time-consuming for identifying patients with specific clinical criteria.
• Natural Language Processing (NLP) and machine learning (ML) offer automated solutions for analyzing large volumes of clinical text.
• Accurate patient phenotyping is crucial for clinical research and patient care.

Purpose of the Study:

• To evaluate traditional NLP and ML classifiers for identifying Systemic Lupus Erythematosus (SLE) patients.
• To compare the performance of established methods with a novel Bayesian word vector approach (Word2Vec inversion).
• To determine the most effective computational methods for SLE phenotyping from clinical notes.

Main Methods:

• Utilized clinical notes from 662 SLE patients and controls.
• Generated feature matrices using Bag-of-Words (BOWs) and Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs).
• Applied various NLP classifiers including neural networks, random forests, Naïve Bayes, support vector machines, and Word2Vec inversion.

Main Results:

• Random forests with BOWs achieved 95.25% accuracy and 0.994 AUC.
• Shallow neural networks with CUIs reached 92.10% accuracy and 0.970 AUC.
• Word2Vec inversion showed 90.03% accuracy and 0.905 AUC, not significantly outperforming baseline ICD-9 codes.

Conclusions:

• Shallow neural networks (with CUIs) and random forests (with CUIs and BOWs) are the most effective classifiers for SLE phenotyping.
• While Word2Vec inversion did not outperform other methods in this study, it shows promise due to its feature-less nature and adaptability.
• Further data may enhance the performance of the Word2Vec inversion method for future applications.