Personalized survival benefit estimation from living donor liver transplantation with a novel machine learning method
Anirudh Gangadhar1, Bima J Hasjim2, Xun Zhao3
1Transplant AI Initiative, Ajmera Transplant Centre, University Health Network, University of Toronto, ON, Canada; Toronto General Hospital Research Institute, University Health Network, Toronto, ON, Canada.
Background & Aims:
Addressing many clinical questions, such as estimating survival differences between living donor (LDLT) and deceased donor liver transplantation (DDLT), relies on observational studies, as randomized-controlled trials (RCTs) are often unfeasible. Thus, we developed decision path similarity matching (DPSM) - a novel machine learning (ML)-based algorithm that simulates RCT-like conditions to mitigate confounding in observational data.
Methods:
We conducted a retrospective study of adult (≥18-years-old) LT candidates between 2002-2023 using the Scientific Registry of Transplant Recipients database. A random forest classifier was trained to predict transplant type from clinicodemographic characteristics. After hyperparameter tuning, decision paths were extracted for individual patients and tree-averaged Hamming distances (dh) were computed for every LDLT-DDLT decision path pair. One-to-one matching was performed by minimizing the total dh across all patient pairs. Random survival forest models were then trained on the matched cohorts to predict post-transplant survival.
Results:
Of 72,581 LT recipients, 93.8% underwent DDLT and 6.2% underwent LDLT. After matching LDLT with DDLT recipients, DPSM successfully reduced confounding associations as shown by a decrease in AUROCpost-match from 0.82 to 0.51. Random survival forest models outperformed traditional Cox regression in both groups (C-indexldlt 0.67 vs. 0.57; C-indexddlt 0.74 vs. 0.65). The predicted 10-year mean survival gain for LDLT over DDLT was 10.3% (SD = 5.7%). In particular, the survival benefit from LDLT was greatest for primary sclerosing cholangitis (12.4% ± 5.3%) and HCV (12.1% ± 4.7%) compared to other etiologies.
Conclusions:
DPSM offers a novel ML-based method for simulating RCT-like conditions in observational data, enabling personalized survival prediction while minimizing confounding. This approach equips clinicians with a new tool to more confidently evaluate treatment effects.
Impact And Implications:
Living donor liver transplantation (LDLT) has emerged as an effective strategy to expand the donor pool, though data from randomized-controlled trials (RCTs) are lacking due to ethical and practical barriers. We developed a novel machine learning-based algorithm termed decision path similarity matching (DPSM), which more effectively reduces bias in observational data by creating cohorts that better approximate those in RCTs. Using DPSM, LDLT was associated with a predicted 10-year mean survival gain of 10.3% (SD = 5.7%) over deceased donor liver transplantation. LDLT was also shown to be particularly beneficial for certain etiologies, i.e. HCV and PSC. DPSM provides clinicians with a powerful tool that transforms real-world observational data into an RCT-like framework, making it an invaluable method in situations where true randomization is not feasible.
