Examining the feasibility of data-driven decision support for the virtual crossmatch for solid organ transplantation: A single center study

The virtual crossmatch is used in transplant medicine to assess the compatibility of organ donors and recipients. Virtual crossmatch methods vary considerably across institutions; require highly trained HLA laboratory experts and clinicians for interpretation; and do not generate data in a standardized format suitable for comparison across institutions. It is not known if standardized multi-center data collection and reporting could potentially facilitate the development of data-driven immunologic decision-making. We sought to examine the feasibility of an algorithmic approach to interpreting virtual crossmatch data.

We examined Histocompatibility and Immunogenetics laboratory data from 1,152 transplant patients and 1,180 donors from an academic medical transplant center over a ten-year time interval. Principal component analysis was used to simplify the complex high-dimensional data with rare outcomes into a format better suited for analysis. Machine learning models were used to predict negative flow crossmatch results. A training subset of the oldest 80% of the data was used to identify the top-performing model. The model's performance was assessed on the newest 20% of the data with the area under receiver operating characteristic curve (AUC).

The final dataset included 2205 crossmatch results from 1446 patient-donor pairs of which 2019 (91.6%) were negative and 186 (8.4%) positive. The top-performing model test set AUC was 0.80.

This study offers the first proof-of-concept of the feasibility of an algorithmic approach to estimate physical crossmatch results. Standardized, multi-institution data collection is necessary to further explore the possibility of a standardized, data-driven virtual crossmatch process.