BC Cancer Registry trains tumor classifiers on unlabeled pathology reports

The paper's actual contribution is methodological: showing that attention-based MIL can extract document-level signal from aggregate patient outcomes without manual annotation, which is distinct from simply applying existing MIL to a new domain. The constraint that makes this work is the operational reality of cancer registries, not the algorithm itself.

This connects directly to the weak-supervision pattern we covered in the Dynamic Bidirectional Pattern Memory study (early July). That work showed learned gating rules fail at scale in clinical NLP when failure modes fragment across rare variants, forcing practitioners toward static, interpretable alternatives. The BC Cancer Registry approach sidesteps this by working with coarse labels that already exist operationally, rather than trying to learn fine-grained rules from sparse rejection logs. Both papers expose the same constraint: clinical ML systems must work within the supervision budget that institutions actually have, not the ideal budget researchers wish they had.

If BC Cancer Registry publishes external validation results on a held-out hospital system within 12 months, that confirms the MIL framework generalizes across institutions with different tumor coding practices. If the approach remains internal-only or shows significant performance drops on external data, it signals the method is tuned to BC's specific registry structure rather than solving the general weak-supervision problem.