Secondary Logo

Journal Logo

Is it Still Tough to Make Predictions About the Future?

Emond, Jean C. MD1

doi: 10.1097/TP.0000000000002839
Commentaries
Free

1 Department of Surgery, Columbia University, New York City, NY.

Received 6 June 2019.

Accepted 11 June 2019.

Correspondence: Jean C. Emond, MD, Thomas S. Zimmer Professor of Surgery, Chief of Transplantation Services, Columbia University and The New York Presbyterian Hospital, 622 W 168th St, New York, NY 10032. (je111@cumc.columbia.edu).

In the current issue of Transplantation, Molinari et al1 present a provocative argument that post-liver transplant survival is nearly exclusively dependent on pretransplant recipient variables. The authors performed an analysis of a large UNOS data set and demonstrated that post-transplant mortality was predicted with a high accuracy using variables available at the initial evaluation of the candidate. The math and analysis informing this project are clearly impressive, particularly using the leverage provided by machine learning to discern predictive relationships hidden in the mass of data from 30 458 patients. The obvious questions that arise from this analysis are, first, is this real? And, second, how will this inform clinical practice?

The C-statistic of 0.952 reported by the authors exceeds the predictive power of all the predictive models in actively used in transplantation, including the model for end-stage liver disease,2 the donor risk index,3 the UK donation after circulatory death risk score,4 and many others. The elaborate risk adjustment models used by the Scientific Registry of Transplant Recipients for transplant oversight by the Organ Procurement and Transplant Network and Centers for Medicare & Medicaid Services could hardly perform better.5 This raises several possible explanations, first, that this team has discovered a collection of variables is better than any other combination in repeated explorations of the United Network for Organ Sharing data over the past 2 decades. While exciting, it is more likely that the machine learning algorithm is more effective at significance mining than traditional manual, or hypothesis-driven exploration of the data sets. Has the machine finally discovered elements that have previously eluded us?

The second challenge posed by this study is the argument that only preoperative data are needed to accurately predict posttransplant survival. The results certainly speak for themselves, and at first glance, it appears that this is indeed the case. Several past efforts such as the Survival Outcomes Following Liver Transplantation score6 have attempted to combine donor and recipient data to help surgeons select donors for recipients to optimize transplant outcomes. The practical value of such a model is obvious, yet there is little evidence that such scores are often used in daily practice. The low predictive power of donor risk index3 is certainly supportive of Molinari et al1 contention that donor variables add little to the value of their predictive model.

The fundamental challenge faced by model builders who use UNOS and other registry data is that only donors and recipients that are actually used for transplantation are available for study. Donors and recipients with clinical criteria deemed too risky for transplantation are excluded from the data set by decisions made every day in clinical practice. It has been noted that the large number of discarded organs have clinical features that overlap with donors that are actually used.7

A 5-year patient survival exceeding 50% is commonly accepted as the floor of ethical selection of liver transplant recipients avoiding futile transplants. The author’s data are consistent with this in that even the worst risk group are observed to have survival exceeding this threshold. Regardless of the accuracy of the predictive score, the utility of the model at the bedside remains in question. It is certainly beneficial to accurately counsel patients when obtaining informed consent for surgical procedures. However, the model does not identify futile transplants since there are few such transplants in the UNOS data based in deliberate selection in practice and so no patients would be turned down for transplantation even with the worst score.

In summary, we once again identify the parameters of clinical practice that are associated with good results. Beyond this, our mandate to push the boundaries of donor and recipient selection to save more lives through transplantation remains unchanged. Constantly improving predictive models should guide the way forward.

Back to Top | Article Outline

REFERENCES

1. Molinari M, Ayloo S, Tsung A, et al. Prediction of perioperative mortality of cadaveric liver transplant recipients during their evaluations.Transplantation2019103e297–e307
2. Wiesner R, Edwards E, Freeman R, et al; United Network for Organ Sharing Liver Disease Severity Score CommitteeModel for end-stage liver disease (MELD) and allocation of donor livers.Gastroenterology200312491–96
3. Feng S, Goodrich NP, Bragg-Gresham JL, et al. Characteristics associated with liver graft failure: the concept of a donor risk index.Am J Transplant20066783–790
4. Schlegel A, Kalisvaart M, Scalera I, et al. The UK DCD risk score: a new proposal to define futility in donation-after-circulatory-death liver transplantation.J Hepatol201868456–464
5. Wey A, Salkowski N, Kasiske BL, et al. Comparing scientific registry of transplant recipients posttransplant program-specific outcome ratings at listing with subsequent recipient outcomes after transplant.Am J Transplant201919391–398
6. Rana A, Hardy MA, Halazun KJ, et al. Survival outcomes following liver transplantation (SOFT) score: a novel method to predict patient survival following liver transplantation.Am J Transplant200882537–2546
7. Wey A, Salkowski N, Kasiske BL, et al. Influence of kidney offer acceptance behavior on metrics of allocation efficiency.Clin Transplant201731e13057
Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.