Probability calibration on the example of improving early cancer detection: a fuzzy set theory approach

Cancer is the leading cause of death before the age of 70 years. Cancer mortality is reduced by early detection. To improve the early diagnosis of cancer, we propose a novel probability calibration method based on a fuzzy set theory. Our model for binary classification was tested on the detection of female breast cancer and lung cancer. The first case is complicated by a small data set problem, while the second – by highly imbalanced data. In both cases, our probability calibration method improved Log Loss (the best result is improved on 48.86%), Brier score (the best result is improved on 13.24%), and the area under the PR curve (the best result is improved on 13.94%). The application area of our algorithm can be extended to any progressive diseases and events without a clearly defined boundary.