Korean J Transplant.  2022 Nov;36(Supple 1):S137. 10.4285/ATW2022.F-2590.

Development of Korean lung allocation system using machine learning

• Yeo HJ
• Jang JH
• Cho WH
• Kim D
• Kim T

• KMID: 2535653
• DOI: http://doi.org/10.4285/ATW2022.F-2590

Abstract

Background
The shortage of donor lungs in Korea has raised ethical demands to optimize organ allocation. We developed a Ko-rean lung allocation system (LAS) that maximizes the transplant benefit using the Korean Network for Organ Sharing data.
Methods
Transplant benefit was defined as a high probability of dying within 1 year of the waiting list and a high probability of surviving at 1 year after transplantation. From 2010 to 2020, 1,587 registered patients for lung transplantation aged 12 years and older and 760 lung transplant patients in Korea were included in the analysis. Through elastic net Cox regression, each model was created to predict death within 1 year on waitlist, and 1 year death after transplantation, and the two models were combined. The final model was validated through the validation cohort and compared with LAS score in US.
Results
The waitlist mortality model included hospitalization at registration, ventilator, extracorporeal membrane oxygenation, gender, age, body mass index, first status at registration, diagnosis, and blood type, and the C-index of training cohorts was 0.801 (P<0.001) and C-index of test cohorts was 0.858 (P<0.001). The transplant mortality model included hospitalization at registration, ventilator, age, body mass index, first status at registration, status at transplantation, diagnosis, and blood type, and the C-index of training cohorts was 0.645 (P<0.001) and C-index of test cohorts was 0.814 (P<0.001). In the weighted sum model, AUC was 0.655 in training cohorts (P<0.001) and AUC was 0.630 in test cohorts (P<0.001).
Conclusions
Compared to the existing LAS score in US, the newly developed Korean LAS showed better performance for pre-dicting transplant benefit. Prospective validation of this model and further refinement of the model are needed.