Global Explainability of A Deep Abstaining Classifier for Cancer Pathology Reports

We present a global explainability method to characterize sources of errors in a real-world multitask deep abstaining classifier (DAC), in the context of cancer histology prediction. Our multitask classifier, currently deployed for automated annotation of cancer pathology reports from NCI-SEER registries, was trained and evaluated on 1.04 million hand-annotated samples and makes simultaneous predictions of cancer site, subsite, histology, laterality, and behavior for each report. The DAC framework enables the model to abstain on ambiguous reports and confusing classes to achieve the target accuracy on the retained (non-abstained) samples, but at the cost of decreased coverage. Requiring 97% accuracy on the histology task caused our model to retain only 22% of all samples, mostly the less ambiguous and common classes. Local explainability with the GradInp technique provided a computationally efficient way of obtaining contextual reasoning for hundreds of thousands of individual predictions. Our method, involving dimensionality reduction of approximately 13000 aggregated local explanations (ALE), offers a tractable path to true global explainability. It enabled identification of sources of errors in histology classification, globally, as hierarchical complexity among classes, label noise, insufficient information, and conflicting evidence. This suggests several strategies for iterative improvement of our DAC, including well-designed exclusion criteria, focused annotation, and reduced penalties for errors involving hierarchically related classes.