Multitask Learning with BERT for Automated Completion of the Breast Cancer Pathology COSD

Breast cancer pathology reporting is crucial for clinical decision-making and national cancer registration in the UK, and it contributes to the Cancer Outcomes and Services Dataset (COSD). Currently, manual extraction of key variables from free-text reports causes data quality issues and delays. This thesis addresses these challenges by developing a novel Multitask BERT (MT-BERT) framework for automated extraction of structured COSD breast pathology attributes. The proposed model combines a BioMedBERT encoder with specialised classification and regression heads to predict seven variables, including histology, grade, and continuous tumour size. To tackle severe class imbalance and missing labels in a real-world dataset of 7673 reports for 2455 breast cancer patients, the model employs a Round-Robin head-balanced sampling method and a sophisticated loss function. This combines Class-Balanced Cross-Entropy with Label-Distribution-Aware Margins (LDAM) and a center Logit Adjustment technique specifically designed for rare histological subtypes. Evaluated on an independent test set, the MT-BERT achieved a mean Macro-F1 score of 0.5566 across all classification tasks, with laterality reaching production-ready accuracy and other variables, including grade, showing promise, though not yet meeting clinical decision-support thresholds. The performance on minority classes in Histology Coded reached a Macro-F1 of 0.5179, partially validating anti-collapse strategies. For the continuous Tumour Size variable, predictions yielded a MAE of 12.23 mm on the original scale. The framework demonstrates that a unified, data-aware multitask transformer can effectively replicate structured COSD reporting, offering a parameter-efficient approach with a single shared encoder in place of seven separate models to support data completeness within the NHS.