Using case-level context to classify cancer pathology reports.

Area of Science:

• Medical Informatics
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Individual electronic health records (EHRs) and clinical reports are often sequential, requiring analysis of aggregate case-level data.
• Extracting comprehensive cancer case information necessitates integrating data from multiple reports over a disease's trajectory.

Purpose of the Study:

• To develop a modular add-on for deep learning models to capture case-level context from sequential clinical reports.
• To enhance the accuracy of text classification tasks on cancer pathology reports by incorporating holistic case information.

Main Methods:

• A modular add-on was developed for compatibility with existing deep learning text classification architectures.
• The approach was evaluated on a large corpus of 431,433 cancer pathology reports.
• Classification accuracy was assessed across six key tasks: site, subsite, laterality, histology, behavior, and grade.

Main Results:

• Incorporating case-level context significantly improved classification accuracy across all six evaluated tasks.
• The proposed add-on demonstrated a substantial boost in performance for analyzing cancer pathology reports.
• The method proved effective in capturing aggregate information from sequential EHR data.

Conclusions:

• The developed add-on effectively captures case-level context, significantly enhancing classification accuracy in cancer pathology reports.
• This modular approach is adaptable to various deep learning architectures and clinical text-based tasks.
• Integrating sequential report data improves the understanding and analysis of complex diseases like cancer.