Challenges

Organ non-use rates continue to rise, with kidney non-use reaching 34% nationally, as of 2024. At the same time, longstanding inequities persist in organ allocation and acceptance, particularly affecting women and patients with primary biliary cholangitis (PBC) or primary sclerosing cholangitis (PSC). The field also faces difficulty in objectively identifying organ suitability and recipient match at both the OPO and transplant center levels.

Objectives of the Clinical Interventions

• To build AI tools that can accurately identify organs at risk of non-use and match them with centers most likely to accept them
• To improve fairness in organ allocation by considering real-time patient data and individualized risk
• To provide transplant teams with clear, explainable AI support to enhance clinical decision-making
• To ensure responsible use of AI through transparency, fairness, and collaboration across disciplines

Approaches and Actions Taken

• Dr. Threlkeld introduced a national AI model that links donor, recipient, and transplant center data to predict when a kidney will be accepted or at risk of non-use. Early adoption across several OPOs was made possible through input and collaboration from key stakeholders.
• Dr. Canfield’s team developed three AI models to support different stages of the transplant process—allocation, provisional acceptance, and final acceptance. The project focuses on building trust, reducing bias, and encouraging collaboration between humans and AI. Ongoing research includes testing these models in real-time decision-making using simulated field trials on UNOS’s SimUNet platform.
• Dr. Bhat shared two AI-driven liver transplant innovations:
  • One uses GPT-4 agents as virtual members to simulate a transplant selection committee. Tasks consist of evaluating candidate cases and reaching consensus decisions.
  • The other applies a Dynamic AI Model to better predict waitlist dropout risk and address disparities, particularly among women and patients with PBC/PSC. The model has been silently trialed within Epic for potential real-world integration.

Findings (Data and Results)

• Valiant AI’s implementation led to a 19% increase in transplant success for high-KDPI kidneys within 90 days.
• Experimental data from Dr. Canfield’s work shows that users can better interpret AI output when provided with model limitations and uncertainty levels.
• Dr. Bhat’s dynamic AI model demonstrated a C-index of 0.82 for predicting waitlist outcomes and improved fairness across patient subgroups.

Conclusion

​​Artificial intelligence holds significant promise for improving organ allocation and acceptance, but its impact relies on thoughtful implementation, transparency, and continuous oversight. Presenters emphasized the importance of ongoing evaluation, human-in-the-loop design, and planned clinical trials to ensure fairness, accuracy, and trust. With careful integration, these tools can enhance both outcomes and fairness for patients nationwide.