Gopika Punchhi’s bio

Gopika Punchhi is a third-year medical student at the Schulich School of Medicine and Dentistry, University of Western Ontario. She completed her undergraduate studies at the Johns Hopkins University, Baltimore, MD. She is a member of Dr. Mamatha Bhat’s lab at the Multi Organ Transplant Program at University Health Network, Toronto, ON. Her research interests are primarily in patient selection, organ allocation, and post-transplant outcomes for liver transplantation.

Deep Learning to Predict Trajectories and Identify Features Associated with Death and Transplant in Waitlisted NASH Patients

Gopika Punchhi, Yingji Sun, Sirisha Rambhatla, Mamatha Bhat

Background: Non-alcoholic steatohepatitis (NASH) cirrhosis patients waitlisted for liver transplantation (LT) are older, have more comorbidities, and have a high risk of dropout and death on the waitlist. We used data at time of waitlisting to construct a DeepHit machine learning model and conduct a competing risk analysis to predict the probability of death versus transplant in waitlisted NASH candidates.

Methods: We included 17,551 NASH patients listed for LT from 2002-2021 from the Scientific Registry of Transplant Recipients (excluding exception indications such as hepatocellular carcinoma). We constructed a DeepHit model with death on the waitlist as the primary event and LT as the competing risk and compared it to regularized CoxPH models predicting death or transplant. We selected the best-performing models and hyperparameters based on the average concordance index (C-index) during validation. Event-specific C-indexes and Brier scores were evaluated at the 25th percentile (1 month), median (5 months), and 12 months on the waitlist using the test set and compared between the models.

Results: The DeepHit model achieved greater C-index scores for the death event at 1, 5, and 12 months of 0.920(⁤±0.001), 0.820(±0.002), and 0.750(±0.002), respectively, and achieved smaller Brier scores for the transplant event of 0.130(±0.003), 0.210(±0.001) and 0.220(±0.005), respectively. In the DeepHit model, features associated with death were initial MELD, hospitalization and ICU status, and poor functional status, while initial MELD, AB blood type, poor functional status, and hemodialysis were associated with transplant.

Conclusion: Our DeepHit model outperforms traditional CoxPH models and can be used to predict the trajectory of a NASH patient on the LT waitlist using data at waitlisting. Modifiable predictors of death or transplant can be identified to reduce the risk of waitlist death.