Autoimmune hemolytic anemia with gel-based immunohematology tests: neural network analysis

Abstract

In a previous report, we investigated the capability of commercially available immunohematology tests based on gel technology to add useful information for the diagnosis of autoimmune hemolytic anemia (AIHA). In this report, we analyzed the same casuistic to find useful information on the importance of different immunohematology tests for the AIHA diagnosis, but using the artificial neural network (ANN) analysis. We studied 588 samples with a positive direct antiglobulin test (DAT), of which 52 samples came from patients with AIHA. The samples were analyzed with the ANN using the multilayer perceptron with the backpropagation algorithm. Using the ANN in the observed data set, the predictive value for the presence of AIHAs was 94.7%. The rate of DAT-positive cases that were not AIHA and that were correctly classified was 99.4%. The receiver operating curve area for the model was 0.99. The independent variable importance analysis found that the gel centrifugation test anti-IgG titer was an important contributor to the network performance, but other variables such as the IgG subclasses can also be considered important. The use of the ANN permitted us to identify immunohematology tests that were "hidden" with the common statistical models used previously. This was the case for the IgG subclasses. However, it is very likely that the information given to the network from those tests is quantitative rather than qualitative.