Hierarchical Temporal Attention Networks for Cancer Registry Abstraction: Leveraging Longitudinal Clinical Data With...

Area of Science:

• Medical Informatics
• Computational Oncology
• Health Data Science

Background:

• Cancer registration is essential for public health policy and cancer control.
• Extracting structured data from unstructured clinical reports for cancer registries is complex and time-consuming.
• Existing methods struggle with the nuances of longitudinal patient data.

Purpose of the Study:

• To develop an automated method for cancer registry abstraction from unstructured clinical text.
• To improve the accuracy and efficiency of extracting structured data from longitudinal patient records.
• To enhance the interpretability of automated data extraction for cancer registrars.

Main Methods:

• Proposed a hierarchical temporal attention network with word, sentence, and document-level attention.
• Incorporated temporal and report type information to analyze longitudinal patient data.
• Developed a stratified sampling algorithm for balanced dataset evaluation across 23 coding tasks.

Main Results:

• Achieved an average F1-score of 0.82, outperforming existing approaches.
• Attention mechanisms effectively prioritized task-relevant information.
• Ablation studies confirmed the importance of proposed network components.

Conclusions:

• The hierarchical temporal attention network significantly advances automated cancer registry abstraction.
• The developed visualization tool enhances interpretability for cancer registrars.
• This approach represents a substantial step towards automating the extraction of cancer data from clinical text.