Automatic information extraction from childhood cancer pathology reports

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• The International Classification of Childhood Cancer (ICCC) is crucial for classifying pediatric cancers.
• Machine learning models for ICCC classification have not been previously developed.
• Existing models use ICD-O-3 coding, necessitating extension for ICCC.

Purpose of the Study:

• To develop and evaluate deep learning models for automated ICCC classification from pathology reports.
• To compare direct ICCC classification with recoding from ICD-O-3.
• To assess model performance across different training data sizes.

Main Methods:

• Developed two deep learning models: one for ICD-O-3 recoding to ICCC (Model 1) and one for direct ICCC classification (Model 2).
• Evaluated models on 29,206 pathology reports from 6 state cancer registries (ages 0-19).
• Utilized uncertainty quantification to assess model confidence.

Main Results:

• Direct ICCC classification (Model 2) significantly outperformed the ICD-O-3 recoding model (Model 1).
• Model 2 achieved a micro-F1 score of 0.987 with uncertainty quantification.
• The model abstained from coding only 14.8% of ambiguous reports.

Conclusions:

• Machine learning-based information extraction is a reliable method for classifying childhood cancer pathology reports.
• These models can supplement human annotators at cancer registries.
• Automated extraction accurately and reliably processes the majority of childhood cancer pathology reports.