PASCAL: a pseudo cascade learning framework for breast cancer treatment entity normalization in Chinese clinical text

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Oncology

Background:

• Clinical text data from electronic health records presents challenges due to author- and domain-specific variations.
• Accurate normalization of breast cancer treatment entities is crucial for downstream clinical studies like careflow mining and therapy analysis.
• Existing methods often use pipeline approaches, leading to error propagation, and few address normalization in complex Chinese clinical text.

Purpose of the Study:

• To propose PASCAL, an end-to-end framework for accurate breast cancer treatment entity normalization (TEN).
• To address limitations of existing pipeline methods and the complexity of Chinese clinical text.
• To enhance knowledge discovery from breast cancer treatment records.

Main Methods:

• PASCAL utilizes a gated convolutional neural network for contextual feature extraction and long-term dependency modeling.
• Treatment entity recognition (TER) is incorporated as an auxiliary task for regularization and to enhance the primary TEN task.
• A conditional random field (CRF) layer models the normalization sequence by concatenating context-aware and probabilistic distribution vectors.

Main Results:

• The PASCAL framework was evaluated against three state-of-the-art sequential models on a real-world database.
• Experiments were conducted in both single- and multi-task learning settings.
• PASCAL demonstrated superior accuracy and efficiency compared to existing approaches.

Conclusions:

• The PASCAL framework effectively normalizes breast cancer treatment entities in clinical text, validating its pseudo-cascade learning approach.
• The framework's ability to extract valuable information from unstructured text significantly benefits downstream tasks.
• PASCAL contributes to advancing clinical decision support systems, including treatment recommendations, breast cancer staging, and careflow mining.