Comparison between an artificial neural network and logistic regression in predicting acute graft-vs-host disease after unrelated donor hematopoietic stem cell transplantation in thalassemia patients

Objective

There is growing interest in the development of prognostic models for predicting the occurrence of acute graft-vs-host disease (aGVHD) after unrelated donor hematopoietic stem cell transplantation. A high number of variables have been shown to play a role in aGVHD, but the search for a predictive algorithm is still ongoing. Artificial neural networks (ANNs) represent an attractive alternative to multivariate analysis for clinical prognosis. So far, no reports have investigated the ability of ANNs in predicting HSCT outcome.

Materials and Methods

We compared the prognostic performance of ANNs with that of logistic regression (LR) in 78 β-thalassemia major patients given unrelated donor hematopoietic stem cell transplantation. Twenty-four independent variables were analyzed for their potential impact on outcomes.

Results

Twenty-six patients (33.3%) developed grade II to IV aGVHD. In multivariate analysis, homozygosity for donor KIR haplotype A (p = 0.03), donor age (p = 0.05), and donor homozygosity for the deletion of the human leukocyte antigen−G 14-bp polymorphism (p = 0.05) were independently significantly correlated to aGVHD. The mean sensitivity of LR and ANNs (capability of predicting aGVHD in patients who developed aGVHD) in test datasets was 21.7% and 83.3%, respectively (p < 0.001); the mean specificity (capability of predicting absence of aGVHD in patients who did not develop aGVHD) was 80.5% and 90.1%, respectively (p = NS).

Conclusion

Although ANNs are unable to calculate the weight of single variables on outcomes, they were found to have a better performance than LR. A combination of these two methods could be more efficient in predicting outcomes and help tailor GVHD prophylaxis regimens according to the predicted risk of each patient. Whether ANN technology will provide better predictive performance when applied to other datasets remains to be confirmed.