Personalized survival benefit estimation from living donor liver transplantation with a novel machine learning method for confounding adjustment

Background & Aims

Many clinical questions, such as estimating survival differences between living donor (LDLT) and deceased donor liver transplantation (DDLT), are limited to observational studies and cannot be answered through randomized controlled trials (RCTs). Thus, we developed Decision Path Similarity Matching (DPSM), a novel machine learning (ML)-based algorithm that simulates RCT-like conditions to mitigate confounders in observational data.

Methods

We conducted a retrospective study of adult (≥18-years-old) LT candidates between 2002-2023 from the Scientific Registry of Transplant Recipients (SRTR) database. A Random Forest (RF) classifier was trained to predict transplant type from clinicodemographic characteristics. After hyperparameter tuning, decision paths for individual patients were extracted and tree-averaged Hamming distances ($$) were computed for every LDLT-DDLT decision path pair. One-to-one matching was algorithmically performed between LDLT and DDLT patients based on minimizing total $$ across all patient pairs. Random Survival Forest (RSF) models were trained on the matched cohorts to predict post-transplant survival.

Results

Of 72,581 LT recipients, 93.8% were DDLT while 6.2% were LDLT. After matching LDLT with DDLT recipients, DPSM successfully reduced confounding associations as shown by a decrease in AUROC_post-match from 0.82 to 0.51. Subsequently, RSF (C-index_ldlt=0.67, C-index_ddlt=0.74) outperformed the traditional Cox model (C-index_ldlt=0.57, C-index_ddlt=0.65). The predicted 10-year mean survival gain of LDLT over DDLT was 10.3% (SD = 5.7%). Particularly, PSC (12.4±5.3%) and HCV (12.1±4.7%) had the highest survival benefits from LDLT compared to other etiologies.

Conclusions

DPSM provides an effective approach for creating RCT-like comparability from observational data to predict which patients will benefit most from LDLT. This novel ML-based methodology enables personalized survival predictions while minimizing confounders and offers clinicians a new tool to more confidently evaluate treatment effects.