An Interpretable Multimodal Framework for Regional Organ Transplantation Outcomes

Abstract

The demand for kidney transplants continues to outpace supply, with over 89,792 patients on the waitlist as of September 2024, yet only 27,332 transplants performed in 2023 [1], and 28% of recovered kidneys going non-utilized [2]. In this thesis, we highlight the use of large language model (LLM) embeddings combined with structured tabular data to build a predictive classifier that estimates offer outcomes for kidney donor-recipient matches. For each predictive model deployed, we provide further analysis on the interpretability of these black-box models using a custom-designed SHAP analysis framework. Our study focuses on three distinct U.S. regions (Regions 1,2,3) with markedly different demographics and amounts of data on organ acceptances (Region 1: 43,126 offers with 2.19% acceptance rate, Region 2: 394,640 offers with 1.57% acceptance rate, Region 3: 169,342 with 2.23% acceptance rate in years 2016-2019). Among the baseline XGBoost models, Region 3 achieved the highest performance, with a precision-accept score of 0.929 and accuracy of 0.993 in the test data. Building on this strong foundation, the multimodal TabText model in Region 3 achieved the best performance overall, with a precision-accept score of 0.959 and accuracy of 0.993 after fine-tuning for six epochs. Our findings suggest that increasing the number of text features, extending training epochs, and incorporating explicit numerical values led to improved model performance in Region 3. In Regions 1 and 2, the baseline model outperformed the TabText model, suggesting that data sparsity in these regions may have limited the effectiveness of the multimodal approach and that further hyperparameter tuning is needed. We also present several visualization techniques to enhance model interpretability. Specifically, we developed a novel SHAP explainer that illustrates feature interactions between multimodal inputs, including both tabular and textual data. Additionally, we explored methods to identify regions of high and low model fidelity by mapping per-sample prediction errors onto t-SNE embeddings. Overall, this thesis introduces new directions for transplant research in the context of transformer-based models and interpretable AI. Leveraging data-driven decision-support tools and refining allocation policies are essential steps toward addressing the persistent gap between supply and demand in the kidney transplant landscape.