Deep learning for Arabic healthcare: MedicalBot

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Healthcare
• Medical Informatics

Background:

• Remote and automated healthcare consultations have surged post-COVID-19.
• Medical bots offer 24/7 access, reduced wait times, and cost savings.
• Effective medical bots require high-quality, domain-specific training data.

Purpose of the Study:

• To address the lack of a comprehensive Arabic medical question-answering dataset.
• To introduce the MAQA dataset, the largest of its kind for Arabic healthcare.
• To benchmark deep learning models for Arabic medical chatbots.

Main Methods:

• Creation of the MAQA dataset: over 430,000 Arabic healthcare Q&A pairs across 20 specializations.
• Implementation and comparison of three deep learning models: LSTM, Bi-LSTM, and Transformers.
• Evaluation of model performance using cosine similarity and BLEU score.

Main Results:

• The Transformer model significantly outperformed LSTM and Bi-LSTM.
• The Transformer model achieved an average cosine similarity of 80.81%.
• The Transformer model obtained a BLeU score of 58%.

Conclusions:

• The MAQA dataset is a valuable resource for developing Arabic medical bots.
• Transformer models show strong potential for advancing Arabic natural language understanding in healthcare.
• This work facilitates improved AI-driven healthcare solutions for Arabic-speaking populations.