Causal Explanation-Guided Learning for Organ Allocation

Dear Reviewer iTSa,

Thank you for your careful reading of our manuscript. We appreciate your recognition of the motivation behind the work, the clarity of the presentation, and the potential of treating refusal codes as counterfactual cues. Below we respond point‑by‑point to each of your comments and describe the revisions now reflected in the updated paper.

Evaluation limited to synthetic data.

Actions taken: We added semi-synthetic experiments using real‑world 2021-2024 UNOS-PTR liver‑offer data (n ≈ 1.1 M offers, 24k donors, 46k candidates).

We ran additional experiments using real patients $X$ and real organs $O$ with the same synthetic setup presented in Appendix A.3. This way, the real covariate structures of patients and organs are preserved.

Additionally, we ran experiments where every instance (consisting of a patient $X$, organ $O$, outcome $Y$ and refusal reason $R$) is real. To evaluate our models we considered the real dataset as $\mathcal{D}_{obs}$ and constructed the feasible set $\mathcal{D_F}$ by randomly rearranging real organs and real patients. We considered and encoded real refusal reasons regarding the donor age and cold ischemia time.

Actions taken: We added a table that contains all used real features with a short explanation of what they represent in the appendix.

Interactions between the features cited in a contrastive counterfactual explanation. Yes, our approach supports this. This can occur in two cases which are not mutually exclusive: i) there are multiple refusal reasons which involve different features (this occurs in a minority of the cases) and ii) the provided refusal reason directly involves multiple features. In both cases, the counterfactual sampling space $\mathcal{A}_i$ becomes a hyperspace, making the reasons less informative.

Which parts of CLEXNET are novel compared to previous work. We like to point out that current ML-based allocation studies sidestep the complexity of refusals, assuming every offer is accepted [8,6,7,63,64]. while more clinically-oriented studies use models of limited complexity and neglect causal mechanisms [55,54,53,17,18]. In the context of organ offer acceptance, everything we present in our work is novel.

In the context of ML, the main novelty is indeed the causal learning from refusal reasons by sampling counterfactuals while the domain adversarial learning for balancing representations has already been done in previous work [13,8].

In our work, everything which is highlighted in light-blue is novel. This includes:

• Figure 3: the counterfactual sampling parts and the explanation loss.
• Algorithm 1: embedding of reasons, directional augmentation (Eq. 1, 2), forward pass of counterfactual, max and average of counterfactuals, initialize explanation loss, guarded explanation loss (Eq. 7, 8).
• Eq. 7-9 which are the core of our methodology and represent the explanatory loss.

Actions taken: We made sure to clarify in the footnote that everything highlighted in light blue is novel and part of our main contribution.

Your feedback helped us strengthen the empirical evaluation, clarify our contributions. We hope these additions address your concerns and demonstrate the practical utility and novelty of our approach.

Sincerely, Anonymous Authors

[6] Berrevoets, J., Alaa, A., Qian, Z., Jordon, J., Gimson, A. E., & Van Der Schaar, M. (2021, July). Learning queueing policies for organ transplantation allocation using interpretable counterfactual survival analysis. In International Conference on Machine Learning (pp. 792-802). PMLR.

[7] Berrevoets, J., Jarrett, D., Chan, A., & van der Schaar, M. (2023). Allsim: Simulating and benchmarking resource allocation policies in multi-user systems. Advances in Neural Information Processing Systems, 36, 851-866.

[8] Berrevoets, J., Alaa, A., Qian, Z., Jordon, J., Gimson, A. E., & Van Der Schaar, M. (2021, July). Learning queueing policies for organ transplantation allocation using interpretable counterfactual survival analysis. In International Conference on Machine Learning (pp. 792-802). PMLR.

[13] Bica, I., Alaa, A. M., Jordon, J., & Van Der Schaar, M. (2020). Estimating counterfactual treatment outcomes over time through adversarially balanced representations. arXiv preprint arXiv:2002.04083.

[17] de Ferrante, H., Rosmalen, M. D. R. V., Smeulders, B., Spieksma, F. C., & Vogelaar, S. (2024). A discrete event simulator for policy evaluation in liver allocation in Eurotransplant. arXiv preprint arXiv:2410.10840.

[18] de Ferrante, H. C., Goya, R. L., Smeulders, B. M., Spieksma, F. C., & Tieken, I. (2025). The ETKidney simulator: a discrete event simulator to assess the impact of alternative kidney allocation rules in Eurotransplant. arXiv preprint arXiv:2502.15001.

[53] SRTR. KPSAM 2015 User Guide, 2015.

[54] SRTR. TSAM 2015 User Guide, 2015.

[55] SRTR. LSAM 2019 User Guide, 2019

[63] Wood, N. L., Mogul, D. B., Perito, E. R., VanDerwerken, D., Mazariegos, G. V., Hsu, E. K., ... & Gentry, S. E. (2021). Liver simulated allocation model does not effectively predict organ offer decisions for pediatric liver transplant candidates. American Journal of Transplantation, 21(9), 3157-3162.

[64] Xu, C., Alaa, A., Bica, I., Ershoff, B., Cannesson, M., & van der Schaar, M. (2021, March). Learning matching representations for individualized organ transplantation allocation. In International Conference on Artificial Intelligence and Statistics (pp. 2134-2142). PMLR.

Thank you for your thorough response. My main concern, about experiments with only synthetic datasets, has been addressed. I have raised my score accordingly.

Dear Reviewer AejF,

Thank you for taking the time to read our submission and for highlighting both the promise of leveraging refusal‑reason information and the limitations of our current evaluation. Below we respond to your points in detail and describe the concrete steps we have taken in the revised manuscript.

Evaluation limited to synthetic data.

Actions taken: We added semi-synthetic experiments using real‑world 2021-2024 UNOS-PTR liver‑offer data (n ≈ 1.1 M offers, 24k donors, 46k candidates).

We ran additional experiments using real patients $X$ and real organs $O$ with the same synthetic setup presented in Appendix A.3. This way, the real covariate structures of patients and organs are preserved.

Additionally, we ran experiments where every instance (consisting of a patient $X$, organ $O$, outcome $Y$ and refusal reason $R$) is real. To evaluate our models we considered the real dataset as $\mathcal{D}_{obs}$ and constructed the feasible set $\mathcal{D_F}$ by randomly rearranging real organs and real patients. We considered and encoded real refusal reasons regarding the donor age and cold ischemia time.

Actions taken: We added a table that contains all used real features with a short explanation of what they represent in the appendix.

We hope the new empirical evidence and robustness analyses address your concerns and demonstrate that CLEXNET can meaningfully improve organ‑offer acceptance predictions in practice.

Sincerely, Anonymous Authors

Missing refusal reasons. In the UNOS-PTR dataset there is always a reason for a refusal and they are generated directly by clinicians. This is required by the current OPTN policy. We cite from section 18.3 (Recording and Reporting the Outcomes of Organ Offers) of the current OPTN policy:

"The allocating OPO and the transplant hospitals that received organ offers share responsibility for reporting the outcomes of all organ offers. ... The OPO or the OPTN must obtain PTR refusal codes directly from the physician, surgeon, or their designee involved with the potential recipient and not from other personnel."

Thus, missing reasons are not an issue in the organ setting. However, it is possible that reasons cannot be embedded into feature space (which results in the same problem) and, in other settings missing reasons could occur. In our implementation we handled this case by setting the explanatory loss $\mathcal{L}_{EXPL}$ to 0 for those instances (as the reasons are missing or cannot be embedded).

Actions taken: Added an additional if statement in Algorithm 1 to check whether there is a reason and whether it can be embedded.

Have clinicians reviewed the counterfactual directions. We did not let clinicians review counterfactual directions. However, we would like to point out that these directions are directly provided by clinicians (we refer to the above OPTN policy citation for this). But we agree that a clinician validation of generated counterfactual samples (which depend on both the counterfactual directions and the feasible region $\mathcal{F}$) should be crucial and performed before any real world use.

Some refusal categories are more reliably directional than others. Absolutely, some refusal categories are not directional or cannot be embedded in the feature space.

We are grateful for your insights, which prompted substantial new experiments and refinements.

Sincerely, Anonymous Authors

Dear Reviewer QPcZ,

Thank you for your thoughtful and constructive evaluation of our manuscript. We appreciate your recognition of the importance of the problem, of our new “direction‑only” learning setting, and of the way CLEXNET marries adversarial debiasing with contrastive counterfactual supervision. Below we respond to each of your concerns in detail and describe the revisions that are now reflected in the updated manuscript.

Evaluation limited to synthetic data.

Actions taken: We added semi-synthetic experiments using real‑world 2021-2024 UNOS-PTR liver‑offer data (n ≈ 1.1 M offers, 24k donors, 46k candidates).

We ran additional experiments using real patients $X$ and real organs $O$ with the same synthetic setup presented in Appendix A.3. This way, the real covariate structures of patients and organs are preserved.

Additionally, we ran experiments where every instance (consisting of a patient $X$, organ $O$, outcome $Y$ and refusal reason $R$) is real. To evaluate our models we considered the real dataset as $\mathcal{D}_{obs}$ and constructed the feasible set $\mathcal{D_F}$ by randomly rearranging real organs and real patients. We considered and encoded real refusal reasons regarding the donor age and cold ischemia time.

Actions taken: We added a table that contains all used real features with a short explanation of what they represent in the appendix.

Transferability of results. Results on the semi-synthetic evaluation yield similar conclusions. Due to the fundamental problem of causal inference it is impossible to evaluate these models with real-world factual data as what we are testing is performance over counterfactual data generated by deviations from the current organ allocation policy. However, we would like to point out that, under our assumptions, integrating refusal reasons provided by clinicians cannot harm the performance of the model.

Potentially incorrect refusal reasons. In our methodology (specifically Eq. 7) we relax Assumption 1 (which assumes that refusal reasons are correct) by adding uncertainty in the reasons through the tunable parameter $p_{ex}$. One could even further customize the model and choose different values for $p_{ex}$ based on, for example, the transplant center ID, the refusal code itself or other features. In general, we consider noisy or inaccurate refusal reasons as future work where probabilistic embeddings of reasons, both in terms of direction (accuracy) and in certainty (reliability), could offer a solution.

Actions taken: Added noisy and inaccurate refusal reasons to future work section.

Dear Reviewer cg51,

Thank you very much for your careful assessment of our submission. We appreciate your acknowledgement of the problem’s importance and of our formulation and experimental design. Below we respond point‑by‑point to every concern and describe the revisions now incorporated in the manuscript.

Evaluation limited to synthetic data.

Actions taken: We added semi-synthetic experiments using real‑world 2021-2024 UNOS-PTR liver‑offer data (n ≈ 1.1 M offers, 24k donors, 46k candidates).

We ran additional experiments using real patients $X$ and real organs $O$ with the same synthetic setup presented in Appendix A.3. This way, the real covariate structures of patients and organs are preserved.

Additionally, we ran experiments where every instance (consisting of a patient $X$, organ $O$, outcome $Y$ and refusal reason $R$) is real. To evaluate our models we considered the real dataset as $\mathcal{D}_{obs}$ and constructed the feasible set $\mathcal{D_F}$ by randomly rearranging real organs and real patients. We considered and encoded real refusal reasons regarding the donor age and cold ischemia time.

Actions taken: We added a table that contains all used real features with a short explanation of what they represent in the appendix.

Strength of Assumption 1.

Actions taken: We specified the conditions under which Assumption 1 holds. Assumption 1 holds only if i) a refusal reason is given ($r \neq \emptyset$) and ii) the refusal reason can be embedded in the feature space ($\mathcal{M}(r) \neq \emptyset$).

Actions taken: We explicitly specified that in our methodology (specifically Eq. 7) we relax Assumption 1 by introducing the threshold acceptance probability $p_{ex}$ (instead of assuming certain acceptance like in Assumption 1).

We agree that Assumption 1 is quite strong, however, it is relaxed in our methodology by a tunable parameter $p_{ex}$.

In the UNOS-PTR dataset there is always a reason for a refusal and they are generated directly by clinicians. This is required by the current OPTN policy. We cite from section 18.3 (Recording and Reporting the Outcomes of Organ Offers) of the current OPTN policy:

"The allocating OPO and the transplant hospitals that received organ offers share responsibility for reporting the outcomes of all organ offers. ... The OPO or the OPTN must obtain PTR refusal codes directly from the physician, surgeon, or their designee involved with the potential recipient and not from other personnel."

Thus, it is realistic in real world clinical workflows, considering noisy or inaccurate refusal reasons as future work.

Actions taken: Added noisy and inaccurate refusal reasons to future work section.

Comparison to Rule-Based Systems. Almost all existing work on organ offer acceptance is based on (probabilistic) rule-based systems/estimators. We believe that our comparison against the Random Forest benchmark shows that CLEXNET is more suited for a robust generalisation and causal estimations.

We also compared a CLEXNET based organ allocation policy against concurrent rule-based allocation policies:

From these experiments we conclude that by taking acceptance into consideration it is possible to place more organs, sooner and with less offers.

Actions taken: Added a policy-level experiment to compare a Acceptance-based policy against other concurrent policies.

We are grateful for your insightful feedback, which led us to strengthen the manuscript substantially. We hope the additional real‑world evaluation, relaxed assumptions, and rule‑based comparisons address your concerns about CLEXNET.

Sincerely, Anonymous Authors

[33] Kim, W. R., Biggins, S. W., Kremers, W. K., Wiesner, R. H., Kamath, P. S., Benson, J. T., ... & Therneau, T. M. (2008). Hyponatremia and mortality among patients on the liver-transplant waiting list. New England Journal of Medicine, 359(10), 1018-1026

[38] Malinchoc, M., Kamath, P. S., Gordon, F. D., Peine, C. J., Rank, J., & Ter Borg, P. C. (2000). A model to predict poor survival in patients undergoing transjugular intrahepatic portosystemic shunts. Hepatology, 31(4), 864-871.

[42] Neuberger, J., Gimson, A., Davies, M., Akyol, M., O’Grady, J., Burroughs, A., ... & Blood, U. K. (2008). Selection of patients for liver transplantation and allocation of donated livers in the UK. Gut, 57(2), 252-257.