The van der Schaar Lab has developed a machine learning solution that goes beyond the scope of contemporary organ-to-patient matching policies by also accounting for the rarity of organs available for transplantation.
Organ transplantation is an incredibly high stakes, life-and-death clinical practice. Making the decision to give an organ to one patient over another involves various treatment effect estimates, the allocation of scarce resources, and difficult conversations with patients. Giving an organ to one patient could save a life but also put others on the waitlist at risk of further complications. To complicate things further, waitlists are ever-growing and there is a lot of uncertainty around when – if at all – suitable organs will become available.
What is the current state of organ transplantation?
Clinicians around the globe are experiencing increasing pressure to determine the best recipient for each organ that becomes available. Over the last fifteen years, transplantation procedures have doubled in the UK. In Canada, such operations have increased by 42 per cent since 2010 and, according to the Organ Procurement and Transplantation Network, a new patient is added to the transplant waitlist every nine minutes in the United States.
These staggering statistics reveal the need for more holistic organ-to-patient matching policies that focus on both patients and organs to maximise the total life years of the population. Cambridge Centre for AI in Medicine Faculty and researchers from the van der Schaar Lab (led by Mihaela van der Schaar, John Humphrey Plummer Professor of Machine Learning, AI and Medicine at the University of Cambridge), have developed an ingenious – and incredibly practical – machine learning solution that does exactly this.
Enter: OrganITE
In 2018 Prof. van der Schaar developed the Tree of Predictors – a novel algorithm that could accurately predict a patient’s likelihood of surviving after a heart transplant. Discussed in NewsWeek and Technology.org, the algorithm was found to be incredibly promising in its accuracy compared to models currently in use and systems proposed by other research groups.
Despite the advancement of the state-of-the-art, transplantation outcome prediction alone is not enough. Specifically, all contemporary prediction methods work in isolation of the waitlist and require it to be more integrated. Hence, any further advancement requires deep collaboration between machine learning experts, as well as expert clinicians.
Responding to this challenge, the van der Schaar Lab’s dedication to organ transplantation research has culminated in the development of OrganITE – a new individualised treatment effect (ITE) methodology to assist clinicians in making the most optimal decision about organ-to-patient matching.
Published at NeurIPS – the foremost conference on the topic of machine learning research – and produced in collaboration with the UK’s leading Transplant Hepatologist Dr. Alexander Gimson, OrganITE is hugely transformational for healthcare. The first ML system of its kind, OrganITE delivers a greater net life expectancy and prevents as many patient deaths as possible by considering the future availability and distribution of organs themselves. Having easy access to these kinds of predictions will allow practices to maximise the life-saving potential of scarce organs, reduce costs that could be better spent elsewhere, and consider the health of every individual patient on the transplant waitlist before making a decision.
“The uniqueness of organs for transplantation and patients’ varied responses to the organs of different qualities present a tantalising problem for machine learning technology. We have dared to ask: What would happen if we gave this particular organ to this patient on the waitlist?“
Dr. Alexander Gimson, Cambridge University Hospitals NHS Foundation Trust
All in all, a very weighty clinical decision in a highly pressurized environment could be made slightly easier by OrganITE.
Why is allocating organs so difficult?
There are a lot of moving parts involved in organ allocation. These include: patients joining and leaving the waitlist, various organs arriving for transplant, clinicians making decisive life and death compromises, and a policy that decides which patient will receive each available organ. It is a dynamic, time-sensitive process that asks clinicians to consider three very challenging questions:
- How can we assign each organ that becomes available “optimally” to a patient on the waitlist?
- How can we accurately estimate the potential outcomes associated with each patient and each possible organ?
- What is the potential outcome for this case if we don’t allocate the organ to this patient?

Put simply, contemporary organ matching policies are ill-equipped to answer these questions.They tend to be quite rudimentary as they are often solely based on estimated outcomes for the patient in question. In the United States and in many other transplant jurisdictions, organs are predominantly allocated based on the survival probability of the patient without a transplant; i.e. x patient needs it the most so x patient should receive the first organ that has become available. In the UK a slightly different approach is taken whereby the organ is offered to the case with the greatest estimated net life years gained by transplantation. Both these types of policies fail to look at the distribution and availability of future organs, and whether the organ currently available will result in the best possible outcome.
Such policies enable decisions that are based on many assumptions and do not give clinicians the full picture required to make an informed decision – one that maximises the potential of all potential incoming organs for all patients on the waitlist.
These policies ignore the fact that:
- Organs arrive in a stream – While the currently available organ might result in a positive outcome for a patient, future organs could have an even greater positive outcome for them; and
- Patients will die soon if not given an organ – Every patient on the waitlist potentially has access to only a limited number of suitable organs, so an organ’s suitability and availability for every patient needs to be considered.
This is where OrganITE can help!
OrganITE challenges and attempts to rectify the inequity that results from ignoring the above issues. It puts equal value in each year of each patient’s life. So instead of simply optimising life years for the current organ available, OrganITE will assign organs to patients based not only on its own estimates of the potential outcomes but also on organ scarcity. By modelling and accounting for the arrival of new organs, the total life years across the population are significantly increased.
How exactly does OrganITE work?
OrganITE is a prediction-based model made up of two components: a high-dimensional ITE estimator and an organ-to-patient policy. The estimator creates patient-organ pairs from which it builds an outcome predictor. So when a new organ arrives, OrganITE ranks the patients in the wait-list according to three criteria:
- How many years are added to each patient’s life when given the organ, compared to not receiving an organ at all;
- How well the arriving organ compares to each patient’s hypothetical “best” organ; and
- How rare this hypothetical best organ is, for each patient.
This third criteria sets OrganITE apart. Under current queue-based methods, when a patient is in need of a rare type of organ, they might need to settle for a less suitable organ that is available immediately – even if the organ in question may be a better fit for someone else and the next-in-line patient may be in better condition and thus able to wait longer for a second suitable organ to be available. Focusing solely on a patient’s potential outcomes is preventing clinicians from making the most of scarce organs.
OrganITE experiments were conducted on a mixture of real and synthetic data — with the real data consisting of 26 years of British organ transplantation surgery history (approx. 19k patients) — and the results indicated that as much as an additional year of life expectancy was added compared to existing organ-to-patient offering policies. Evidently, prediction-based policies result in better outcomes for all parties involved and ought to be integrated into healthcare.
Clinians remain in control
Machine learning tools like OrganITE have the potential to transform healthcare in the coming years. But we understand that clinicians are worried about giving up control over life-and-death decisions for their patients. OrganITE does not aim to take over the decision-making of clinicians, rather it is meant to be a decision support tool – one that still allows clinicians and their patients to make the final decision on any organ-recipient pairing suggested by the system.
“We want to help clinicians make good decisions. [We ask] what would help a clinician in making a good decision? How would these ML tools work in tandem with clinicians? In the end it’s still the patient and clinician making the decision – OrganITE just provides offers.”
Jeroen Berrevoets, van der Schaar Lab
Replacing the expertise of clinicians is not the intention of machine learning for healthcare – it is instead about supporting clinicians so they can make more informed decisions for their patients.
What to expect in the next few years
Moving forward the Cambridge Centre for AI in Medicine hopes to better understand what aspects of new technologies such as OrganITE and OrganSync can be improved in order to allow clinicians to trust such innovative tools. Trust depends on several factors, including:
- Explainability: On what basis does OrganITE predict the organ being considered will yield x number of years for the patient?
- Interpretability: Why are some patients ranked above others?
- Transparency: Can we easily integrate OrganITE’s components into clinical practice?
- Empirical benefit: What is the accuracy of the predictions provided by OrganITE?
Our researchers intend to demonstrate that tools such as OrganITE can give critical information to clinicians and their patients at the point at which they need to make far-reaching decisions to accept or reject a donor offered by allocation systems. Machine learning is not something to be feared but rather embraced as an additional tool in the clinician’s toolbox.
We are now working on models that can state why a proposed organ is being offered to an individual patient, explain a patient’s risk of dying without a transplant as well as the risk of dying with the proposed organ, and their probability of receiving an offer of similar or better quality in the weeks and months ahead. Easy access to this information will ultimately help clinicians and their patients make more rational choices in highly pressurized transplantation contexts.
——–
Dr. Alexander Gimson is a Physician at Addenbrooke’s Hospital and Consultant Transplant Hepatologist at Cambridge University Hospitals NHS Foundation Trust. Read more about his pioneering research here.
Prof. Mihaela Van der Schaar is Director at the Cambridge Centre for AI in Medicine. Read more about her research at the van der Schaar Lab.
Jeroen Berrevoets is a PhD student in the van der Schaar Lab at the University of Cambridge. Read his publications here.