Association of metastatic pattern in breast cancer with tumor and patient-specific factors: a nationwide autopsy study using artificial intelligence.

Abstract

Background: Metastatic spread is characterized by considerable heterogeneity in most cancers. With increasing treatment options for patients with metastatic disease, there is a need for insight into metastatic patterns of spread in breast cancer patients using large-scale studies.

Methods: Records of 2622 metastatic breast cancer patients who underwent autopsy (1974-2010) were retrieved from the nationwide Dutch pathology databank (PALGA). Natural language processing (NLP) and manual information extraction (IE) were applied to identify the tumors, patient characteristics, and locations of metastases.

Results: The accuracy (0.90) and recall (0.94) of the NLP model outperformed manual IE (on 132 randomly selected patients). Adenocarcinoma no special type more frequently metastasizes to the lung (55.7%) and liver (51.8%), whereas, invasive lobular carcinoma mostly spread to the bone (54.4%) and liver (43.8%), respectively. Patients with tumor grade III had a higher chance of developing bone metastases (61.6%). In a subgroup of patients, we found that ER+/HER2+ patients were more likely to metastasize to the liver and bone, compared to ER-/HER2+ patients.

Conclusion: This is the first large-scale study that demonstrates that artificial intelligence methods are efficient for IE from Dutch databanks. Different histological subtypes show different frequencies and combinations of metastatic sites which may reflect the underlying biology of metastatic breast cancer.