An Improved Double Channel Long Short-Term Memory Model for Medical Text Classification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Chinese medical text classification faces challenges due to complex language segmentation and ambiguous terms, impacting existing algorithm accuracy.
• Word-level neural network models struggle with large datasets of complex medical text, particularly with nuanced terminology.

Purpose of the Study:

• To enhance medical text classification accuracy for improved patient triage and precise treatment.
• To address the limitations of word-level models in handling complex Chinese medical diagnostic terms.

Main Methods:

• An improved Double Channel (DC) mechanism was developed, integrating word-level and character-level embeddings simultaneously within a Long Short-Term Memory (LSTM) framework.
• A hybrid attention mechanism was employed to combine time-step outputs and states, calculating weighted scores for improved generalization.
• Trade-off learning was applied to time-step inputs to enhance the model's learning generalization ability.

Main Results:

• The proposed DC-LSTM model demonstrated significantly superior accuracy and ROC performance compared to the baseline CNN-LSTM model.
• Extensive evaluations were conducted on two distinct datasets, cMedQA and Sentiment140, validating the model's effectiveness.
• The DC mechanism effectively captures complex features in Chinese medical texts, overcoming limitations of traditional word-level models.

Conclusions:

• The DC-LSTM model offers a robust solution for accurate medical text classification, particularly for complex Chinese language data.
• This approach improves the potential for accurate patient triage and subsequent precise treatment strategies.
• The integration of multi-level embeddings and attention mechanisms enhances deep learning model performance in specialized domains.