Prediction of nodal metastasis and prognosis in breast cancer: a neural model.

BACKGROUND

An increasing number of women with breast cancer are detected with the disease at an early stage, when the lymph nodes are not involved. In order to obviate the necessity to carry out axillary dissection, accurate surrogates for lymph node involvement need to be identified. In this paper we have examined the use of a neural network to predict nodal involvement. The neural approach has also been extended to investigate its predictive applicability to the long-term prognosis of patients with breast cancer. A number of established and experimental prognostic markers have been studied in an attempt to accurately predict patient outcome 72 months after first examination.

METHODS

81, unselected patients, presenting clinically, who had all undergone mastectomy for invasive breast carcinoma were considered in this study. A total of 12 markers were analysed for the prediction of lymph node metastasis, while node status itself was used as an additional marker for the prognostic analysis. In this case the outcome related to whether a patient had relapsed within 72 months of diagnosis. In both cases, a number of marker combinations were analysed separately in an attempt to classify those most favourable marker interactions with respect to lymph node prediction and prognosis. Patients were randomly divided into a training set (n = 50) and a test set (n = 31). The simulation was developed using the NeuralWorks Professional II/Plus software (NeuralWare, Pittsburgh, Pa, USA).

RESULTS

In the case of lymph node metastasis, the neural network was able to correctly predict axillary involvement, or otherwise, in 84% of the patients in the test set by considering 9 of the 12 available markers. This represents an improvement of 10% over the traditional approach which considers the tumour grade and size only. The sensitivity and specificity were also shown to be 73% and 90%, respectively. With regard to patient prognosis, again 84% classification accuracy was obtained using a subset of the markers, with a sensitivity of 50% and a specificity of 96%.

CONCLUSIONS

Although this study considered a relatively small sample of patients, nevertheless it demonstrates that artificial neural networks are capable of providing strong indicators for predicting lymph node involvement. There is no longer a need for axillary dissection with all its implications in patient morbidity and demands on clinical resources. The management of breast cancer and the planning of strategies for adjuvant treatments is also facilitated by the use of neural networks for the long-term prognosis of patients.