I worked on developing a machine learning algorithm named FIBERS (Feature Inclusion Bin Evolver for Risk Stratification), designed to aid in kidney transplantation matching. Traditional matching methods focused on antigen-level mismatches in the HLA molecule, overlooking the significant variability at the amino acid level. FIBERS, on the other hand, looks at amino acid mismatches as features to stratify instances into low and high-risk groups.

Initially, FIBERS assumed a fixed risk stratification threshold of 0, wherein instances with 0 feature mismatches were considered low risk, while those with more than 0 were classified as high risk. However, this threshold of 0 was an assumption rather than a holistic data-driven approach. Moreover, HLA allele frequencies vary among ethnicities, resulting in differences in the frequency of low-risk AA MM.

To address these issues, we explored the possibility of implementing an adaptable risk stratification threshold. This approach would enable the machine learning algorithm to learn the optimal threshold value from the data, which might not necessarily be 0. We employed simulated data to test two adaptable risk stratification threshold methods and found that FIBERS consistently achieved high accuracy in stratifying instances based on risk.