Prediction of survival in breast cancer: evaluation of different multivariate models

We compared the performance of multivariate models based on mitotic activity index, lymph node status, and tumor size in the prognostication of breast cancer.

Cox and discriminant models for survival were created for two patient groups: a) 120 breast cancer patients, and b) 86 patients with ductal infiltrating carcinoma. The models were compared with the model of Baak et al (1985).

The models distinguished between dying and surviving patients with an efficiency of 70.9-77.9% in mutual tests. With a single cutoff the model of Baak et al was less efficient (50.8-65.8%). If a region of uncertainty was allowed between two cutpoints, the efficiencies below and above the cutpoints increased. When the uncertain region included a third of the patients, the efficiency varied between 73.8 and 84.7%.

Multivariate models seem to need a region of uncertainty between two discriminating cutpoints. These models resulted in the correct prediction of prognosis in about 75% and more of patients. With different materials the models differed in efficiency. With a region of uncertainty the model of Baak et al performed well with completely independent material.