Tumor registry data were collected between 1994-1996 on 11 primary tumor sites and 15 metastatic sites from 4399 patients. The primary and metastatic sites were cross-tabulated in various ways to identify patterns, and the authors developed algorithms by using multinomial logistic regression analysis to predict the locations of primary tumors based on metastatic patterns.

Three primary tumors had single, dominant metastatic sites: ovary to abdominal cavity (91%), prostate to bone (90%), and pancreas to liver (85%). The liver was the dominant metastatic site for gastrointestinal (GI) primary tumors (71% of patients), whereas bone and lung metastases were noted most frequently in non-GI primary tumors (43% and 29%, respectively).

In a study of combinations of liver, abdominal cavity, and bone metastases, 86% of prostate primary tumors had only bone metastases, 80% of ovarian primary tumors had only abdominal cavity metastases, and 74% of pancreas primary tumors had only liver metastases. A single organ was the dominant source of metastases in 7 sites: axillary lymph node from the breast (97%), intestinal lymph node from the colon (84%), thoracic lymph node from the lung (66%), brain from the lung (64%), mediastinal lymph node from the lung (62%), supraclavicular lymph node from the breast (51%), and adrenal gland from the lung (51%)

CONCLUSIONS: The algorithms that the authors developed achieved a cross-validated accuracy of 64% and an accuracy of 64% on an 1851-patient independent test set, compared with 9% accuracy when a random classifier was used.   see algorithm here