Biomedical semantic indexing by deep neural network with multi-task learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Informatics

Background:

• Biomedical semantic indexing is crucial for bioinformatics and information retrieval.
• Accurate annotation of biomedical citations with Medical Subject Headings (MeSH) is challenging due to imbalanced data.
• Traditional sampling methods are ineffective for semantic indexing with unbalanced datasets.

Purpose of the Study:

• To develop a novel deep learning model for biomedical semantic indexing.
• To address the challenge of imbalanced category distribution in training data.
• To improve the efficiency and accuracy of annotating biomedical citations.

Main Methods:

• A deep serial multi-task learning model was proposed.
• The primary task involved multi-label text classification of biomedical citations, considering label relationships.
• An auxiliary regression task predicted MeSH numbers to accelerate network convergence.

Main Results:

• The proposed model demonstrated superior performance compared to the state-of-the-art MTI solution on the BioASQ-Task5A dataset.
• The model achieved the highest precision among all evaluated solutions.
• Faster convergence speeds were observed compared to naive deep learning methods.

Conclusions:

• The model utilizes a serial, tightly coupled multi-task structure, differing from ordinary parallel structures.
• Satisfactory performance was achieved without reliance on handcrafted features.
• The approach offers an effective solution for biomedical semantic indexing challenges.