Reasoning with uncertainty in pathology: artificial neural networks and logistic regression as tools for prediction of lymph node status in breast cancer patients.

Axillary lymph node status is an important prognostic feature for patients with breast cancer, but the therapeutic value of axillary lymphadenectomy is controversial. It would be useful to be able to predict the status of axillary lymph nodes before lymphadenectomy from prognostic features evaluated in a previous breast biopsy. This prediction would be useful to optimize the treatment of patients with breast cancer who are unlikely to have nodal metastases. We studied 279 patients with invasive breast carcinoma treated with modified radical mastectomy or with lumpectomy combined with axillary lymph node dissection. Prognostic factors evaluated were age, histologic type of invasive tumor, presence of associated ductal and/or lobular carcinoma in situ, lesion size, histologic and nuclear grades, DNA index, presence of multiploidy by flow cytometric analysis, and immunocytochemical expression of estrogen and progesterone receptors, proliferating nuclear cell antigen, and HER-2/neu oncogene. Several probabilistic neural networks (NNs) with genetic algorithms were developed using prognostic features as input neurons and lymph node status (positive or negative) as output neurons. The data were also studied with multiple regression and logistic regression analysis. The best NN model trained with 224 cases using 19 input neurons. It classified correctly 49 (89.0%) of 55 unknown cases (specificity, 97.2%; sensitivity, 80.0%; positive predictive value, 93.8%; negative predictive value, 87.5%). Several statistically significant models could be fitted with both multiple regression and logistic regression. The logistic regression model fitted with 240 cases using 6 independent variables estimated correctly 26 (66%) of 39 holdout cases. NNs and logistic regression models offer potentially useful tools to estimate the status of axillary lymph nodes of breast cancer patients before axillary lymphadenectomy. Future prospective studies with larger groups of patients and perhaps better prognostic markers are needed before these predictive multivariate models become ready for clinical use.