Automatic Classification of Thyroid Findings Using Static and Contextualized Ensemble Natural Language Processing...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Manual classification of electronic medical examination results is time-consuming and error-prone.
• Existing automated methods struggle with nonstandardized data and context.
• Need for accurate automated classification of thyroid conditions from diverse medical records.

Purpose of the Study:

• Develop deep learning models for automatic classification of 3 thyroid conditions: healthy, caution required, and critical.
• Enhance classification accuracy, particularly minimizing false negatives for healthy thyroid predictions.
• Utilize electronic medical records from 284 institutions and open-source data.

Main Methods:

• Proposed Static and Contextualized Ensemble NLP network (SCENT) systems.
• Developed ensemble models using Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and ELECTRA.
• Created two versions: SCENT-v1 (CNN + ELECTRA) and SCENT-v2 (hierarchical CNN + LSTM + ELECTRA).

Main Results:

• SCENT-v1 achieved the highest F1 score (92.56%).
• SCENT-v2 demonstrated high recall (94.44%) and minimized misclassifications for caution-required conditions.
• SCENT-v2 achieved zero classification error for critical conditions when predicting healthy thyroid.

Conclusions:

• The SCENT models show robust classification performance for Korean medical data, addressing data imbalance and language nuances.
• SCENT-v1 highlights the benefit of combining static and contextual token representations.
• SCENT-v2 significantly improves the prediction accuracy of healthy thyroid conditions.