Prediction of liver metastases after gastric cancer resection with the use of learning vector quantization neural networks.

Objective: The aim of our study was to determine whether learning vector quantization neural networks could be used to predict liver metastases after a gastric cancer surgery.

Background: The prediction of tumor recurrence is invaluable for tailoring specific treatment and follow-up strategies for gastric cancer patients. At present, it is still impossible to make reliable predictions of tumor progression. The use of complex mathematical models such as neural networks has already been implemented for the study of various pathophysiological mechanisms, but to date they have never been used for predicting liver metastases after gastric cancer resection.

Methods: A total of 213 patients operated for gastric cancer between 1999 and 2005 were included in our study. They were stratified in a model development (140 patients) and validation group (73 patients). With the use of an auxiliary regression network, seven clinicopathological variables were selected to predict liver metastases.

Results: Forty-one patients developed liver metastases (19.2%). The longest follow-up was 2,754 days. Most liver metastases occurred in the first 799 days after discharge. All predictions were compared to actual recurrences with a two by two contingence table. The determined sensitivity and specificity for the development sample were 71 and 96.1%, respectively. The values for the test sample were 66.7 and 97.1%, respectively. The significance of the model was determined using various post-hoc tests, which all confirmed the effectiveness of our model.

Conclusion: The presented model exhibited a high negative predictive value and reasonable high sensitivity for liver metastases. To improve sensitivity, the inclusion of more patients and perhaps biological markers is still necessary.