A Generic Semi-Supervised and Active Learning Framework for Biomedical Text Classification.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Natural Language Processing

Background:

• Biomedical text classification demands extensive manual labeling by clinical specialists, which is resource-intensive.
• Active learning (AL) aims to minimize labeling costs by selecting informative samples but often underutilizes unlabeled data.
• Semi-supervised learning (SSL) can leverage unlabeled data to enhance classifier performance and data representativeness.

Purpose of the Study:

• To propose a generic semi-supervised learning framework to enhance active learning for biomedical text classification.
• To reduce the number of manually labeled training examples required from clinical specialists.
• To improve the overall performance and efficiency of biomedical text classifiers.

Main Methods:

• The proposed framework integrates active learning with semi-supervised learning techniques.
• It combines manually annotated samples selected by active learning with pseudo-labeled data generated by a classifier.
• The framework was evaluated using three biomedical datasets related to obesity and smoking habits across various classification algorithms.

Main Results:

• The semi-supervised framework successfully reduced the manual labeling effort by 10% across tested datasets.
• Classifier performance was maintained despite the significant reduction in manually labeled training examples.
• The framework's effectiveness was consistent across different classification algorithms used in the evaluation.

Conclusions:

• The proposed semi-supervised learning framework effectively reduces the burden of manual annotation in biomedical text classification.
• This approach offers a practical solution for creating high-performing classifiers with reduced reliance on clinical specialists' time.
• The integration of AL and SSL presents a promising strategy for efficient biomedical text data analysis.