A cost-minimizing diagnostic methodology for discrimination between neoplastic and non-neoplastic brain lesions:...

The purpose of this study was to make an improvement in the performance of a logistic regression model in predicting the presence of brain neoplasia with magnetic resonance spectroscopy data by using a new approach for logistic regression coefficient estimation. This new approach, termed cost minimizing (C-min), introduced by one of the authors (Chetty), uses the cost function for prediction outcomes to estimate model coefficients and the prediction decision rule. To do this requires use of a genetic algorithm.

Consecutive patients with suspected brain neoplasms or recurrent neoplasia referred for magnetic resonance spectroscopy were enrolled once a final diagnosis was established by histopathology or clinical course, laboratory data, and serial imaging. For the same magnetic resonance spectroscopy explanatory (input) variables, logistic regression models were constructed with conventional and C-min coefficient estimates, and sensitivity and specificity outcomes were compared at alternative probability threshold levels.

The C-min approach dominated the conventional approach in 14 of 18 trials, in that C-min had either fewer of both false negatives and false positives, or it had the same number of one type, and less of the other type of diagnostic error. C-min was always less costly.

The C-min approach to logistic or other regression model estimation may be a step forward in reducing the cost and, often, the errors of diagnostic (and treatment) processes. However, this new approach must be validated on larger and more varied datasets, and its statistical performance characteristics determined before it can be implemented as a practical clinical tool.