Deep learning survival analysis for clinical decision support in deceased donor kidney transplantation

deep learning, kidney transplant, survival analysis, machine learning, srtr, mortality

Ruales Rosero, Paul E. ; Koyejo ; Sanmi

关键词: deep learning, kidney transplant, survival analysis, machine learning, srtr, mortality;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/104893/RUALESROSERO-THESIS-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

In deceased donor kidney transplantation, the decision to accept or decline an offer relies on a clinicians intuition and ability to digest complex information in order to maximize patient survival. Risks affecting patient survival post-KT must be balanced with the risks of remaining on the waitlist. These risks include mortality, graft failure, and becoming too sick to transplant. The allocation system today takes these risk into account by way of the KDPI and EPTS scores. While these scores are discriminative of patient survival they were built with an assumption of independence between risks and very few donor-recipient variables. Deep learning survival analysis can effectively handle competing risks and learn complex relationships between many more donor-recipient variables. We used DeepHit to assess the risk benefit associated with accepting a kidney offer or remaining on the waitlist. Our models achieved comparable, if not better performance in certain tasks, with other high performing models in the literature and revealed that decoupling competing risks led to increased clinical information gain. We show that comprehensively modeling competing risks using machine learning can achieve more granular, meaningful clinical risk analysis enabling more effective decision making in deceased donor kidney transplantation.