Background: Liver transplantation (LT) is a life-saving treatment for end-stage liver disease, yet long-term graft survival is threatened by injury from rejection, recurrent metabolic liver disease, biliary complications, infections, and other causes. Graft injury may progress before clinically obvious abnormalities emerge.

Current surveillance relies on laboratory tests, imaging, and liver biopsy. Although biopsy remains the reference standard, it is invasive, cannot be performed frequently, and provides only an episodic snapshot of an evolving biological process. There is a critical need for non-invasive approaches that integrate information across the post-transplant course to identify current graft injury and anticipate deterioration.

Research Aims & Methods: In preliminary work, we developed a clinician–AI hybrid multiclass model using >5,000 liver biopsies to classify graft injury etiologies and a longitudinal deep-learning model to predict significant graft fibrosis. These studies support combining longitudinal transplant data, advanced AI, and clinician expertise.

Our central hypothesis is that signals of graft injury and progression are distributed across the post-transplant trajectory and can be identified by integrating continuous-time modelling, clinician knowledge, adaptive learning, and graft-state dynamics.

• Aim 1: Develop and externally validate an adaptive, clinician-informed longitudinal model for pre-biopsy classification of liver graft health states and injury etiologies. Using ~4,150 contemporary UHN biopsies, including protocol biopsies representing normal graft health, we will integrate serial laboratory, immunosuppressant, immunological, imaging, transplant, donor, and recipient data. A continuous-time Transformer will model irregular trajectories. Predictions will be combined with hepatologist-derived diagnostic knowledge through Bayesian fusion. Bounded continual learning will enable adaptation to evolving clinical practice while limiting catastrophic forgetting. External validation will use cohorts from Mayo Clinic and NUHS Singapore.
• Aim 2: Develop and externally validate a clinically constrained AI world model of liver graft health for continuous state estimation and early forecasting of deterioration. The model will learn a continuously updated latent representation of graft health from longitudinal multimodal data. A state-transition model will learn how this state evolves, while outcome models forecast significant graft injury and fibrosis progression over 3-, 6-, 12-, and 24-month horizons. We will assess discrimination, calibration, uncertainty, state transitions, and lead time relative to conventional clinical triggers.

This research integrates expertise in transplant hepatology, liver pathology, biostatistics, and computer science.

Clinical and pathology expertise will define graft states and outcome anchors; computational expertise will advance continuous-time learning, clinician–AI integration, continual learning, and world modelling. Anticipated Outcomes: This research will establish an externally validated, interpretable framework for proactive liver graft surveillance that identifies the likely cause of current injury and forecasts deterioration. It could enable earlier intervention, more targeted use of biopsy and other investigations, and improved graft preservation. Methodologically, it will advance AI for irregular, partially observed, evolving clinical systems beyond transplantation.