Improved feature selection for hematopoietic cell transplantation outcome prediction using rank aggregation

This paper presents a methodology for developing an improved feature selection technique that will help in accurate prediction of outcomes after hematopoietic stem cell transplantation (HSCT) for patients with acute myelogenous leukaemia (AML). Allogeneic HSCT using related or unrelated donors is the standard treatment for many patients with blood related malignancies who are unlikely to be cured by chemotherapy alone, but survival is limited by treatment-related mortality and relapse. Various genetic factors such as tissue type or human leukocyte antigen (HLA) type and immune cell receptors, including the killer-cell immunoglobulin-like receptor (KIR) family can affect the success or failure of HSCT. In this paper we aim to develop a novel, aggregated ranking based feature selection technique using HLA and KIR genotype data, which can efficiently assist in donor selection before BMT and confer significant survival benefit to the patients. In our approach we use a rank aggregation based feature selection technique for selecting suitable donor genotype characteristics. The result obtained is evaluated with classifiers for prediction accuracy. On average, our algorithm improves the prediction accuracy of the results by 3-4% compared to generic analysis without using feature selection or single feature selections algorithms. Most importantly the selected features completely agree with those obtained using traditional statistical approaches, proving the efficiency and robustness of our technique which has great potential in the medical domain.