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

Cancer registration is a vital source of information for government-driven cancer prevention and control policies. However, cancer registry abstraction is a complex and labor-intensive process, requiring the extraction of structured data from large volumes of unstructured clinical reports. To address these challenges, we propose a hierarchical temporal attention network leveraging attention mechanisms at the word, sentence, and document levels, while incorporating temporal and report type information to capture nuanced relationships within patients' longitudinal data. To ensure robust evaluation, a stratified sampling algorithm was developed to balance the training, validation, and test datasets across 23 coding tasks, mitigating potential biases. The proposed method achieved an average F₁-score of 0.82, outperforming existing approaches by prioritizing task-relevant words, sentences, and reports through its attention mechanism. An ablation study confirmed the critical contributions of the proposed components. Furthermore, a prototype visualization tool was developed to present interpretability, providing cancer registrars with insights into the decision-making process by visualizing attention at multiple levels of granularity. Overall, the proposed methods combined with the interpretability-focused visualization tool, represent a significant step toward automating cancer registry abstraction from unstructured clinical text in longitudinal settings.