Exploring Explanation Multiplicity: Insights from Lung Transplant Survival Data

Both traditional regression-based models and advanced machine learning (ML) approaches are susceptible to model multiplicity, also known as the Rashomon Effect, a term introduced by Breiman (2001). This phenomenon describes the presence of multiple valid explanations (e.g., variable importance or statistical significance) that yield nearly identical predictive performance. While often seen as a challenge, model multiplicity can enhance decision-making by offering flexibility in decision support systems (DSSs). It enables model developers to select or refine models based on desirable properties without compromising prediction accuracy. The existing literature on discovering multiple model explanations is still relatively sparse. Most prior work has concentrated on tree-based models to identify all or part of the Rashomon set, typically relying on greedy or approximate methods that limit exploration of the entire feature space (Xin et al., 2022). In this study, we introduce a practical, efficient, and model-agnostic framework designed to uncover multiple high-performing explanations for classification tasks. Our approach is well-suited for real-world, feature-rich datasets and can identify distinct clusters of important variables within a reasonable computational timeframe. In the healthcare domain, physicians and medical decision-makers often prefer having access to a set of effective models to choose from based on specific clinical contexts, rather than depending on a single model. Model multiplicity enables them to select models that include or exclude particular features of interest for survival analysis and diagnostic purposes, without compromising predictive accuracy. This flexibility is especially valuable in handling missing patient data, as it allows the use of models that rely on the most readily available information. To demonstrate the applicability of the proposed framework, we applied it to a study examining factors influencing short- and long-term survival following lung transplantation. Although advances in organ allocation policies and surgical practices have been made, lung recipients still face lower long-term survival rates compared to recipients of other organs. Leveraging extensive, feature-rich nationwide datasets from the United Network for Organ Sharing (UNOS), the framework systematically explores multiple predictive models that capture key factors, including those related to recipients, donors, and the transplant procedure, influencing lung transplant survival. The insights gained from this study can support policymakers in refining screening, registration, and organ allocation practices and developing more effective indicators for lung allocation, ultimately leading to improved transplantation outcomes.