Medical Text Classification Using Hybrid Deep Learning Models with Multihead Attention

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Automatic medical text classification is crucial for extracting information from clinical descriptions.
• Current machine learning methods require significant human effort for data labeling.
• Electronic health records contain vast amounts of valuable patient data.

Purpose of the Study:

• To propose novel deep learning architectures for medical text classification.
• To reduce the human effort required for creating labeled training data.
• To improve the efficiency of processing detailed patient information in medical texts.

Main Methods:

• Implementation of a quad channel hybrid long short-term memory (QC-LSTM) deep learning model.
• Development and implementation of a hybrid bidirectional gated recurrent unit (BiGRU) deep learning model with multihead attention.
• Validation of proposed models on two distinct medical text datasets.

Main Results:

• The QC-LSTM model achieved a classification accuracy of 96.72%.
• The hybrid BiGRU model achieved a classification accuracy of 95.76%.
• Both models demonstrated high performance in medical text classification tasks.

Conclusions:

• The proposed deep learning models effectively classify medical texts.
• These novel architectures significantly mitigate the need for manual data labeling.
• The developed models offer a promising solution for efficient clinical and translational research.