The question of which patient should receive a life-saving donor organ is one of the thorniest in medicine. Donor organs are a scarce resource, each one as unique as the patients who need a transplant. Who will benefit the most from the transplant? Who will live the longest if they receive the transplant, or die soonest if they do not? If the patient “next in line” for a transplant passes up an offered organ, might a more suitable organ become available in time? Accurately estimating the potential outcomes for each patient and each potential organ is an enormous challenge.
The current approach in the UK to this challenge is to employ mathematical models that seek to maximise the life-years that the currently available donor organ will add to someone’s life as a result of a transplant, judged from the point when they are registered on a transplant list.
Researchers led by Mihaela van der Schaar, the Director of the Cambridge Centre for AI in Medicine, have developed a machine-learning methodology called OrganITE that takes organ-matching technology to a new level. The paper – “OrganITE: Optimal transplant donor organ offering using an individual treatment effect” – will be presented at a premier international ML conference, NeurIPS, in December 2020. It is one of 18 papers accepted at NeurIPS from CCAIM faculty members.
The new system, envisaged as a clinical decision-support tool, recommends an organ be first offered to a selected recipient based not only on its own prediction of the greatest net life years gained but also, crucially, on the likelihood of additional organs becoming available – known as organ scarcity. By combining these two elements, OrganITE has been shown capable of significantly increasing total life years across the patient population as a whole, compared with current approaches. Arguably a better outcome at a societal level.
“Decision-support tools that help patients and their doctors to decide whether they should accept a donor offer – and is capable of explaining to them why it is a rational choice – will be critically important in the future,” says paper co-author Dr Alexander Gimson, Consultant Transplant Hepatologist at Cambridge University Hospitals NHS Foundation Trust and a member of the CCAIM team. Dr Gimson has been a liver transplant physician for over 20 years and is a former chairman of the national organ allocation committee.
Explicitly modelling organ scarcity when considering organ allocation is a critical new development in organ-allocation technology. In this context, scarcity refers not only to absolute numbers of donor organs, but also to their characteristics, and how well they match the organ requirements of patients on the transplant waiting list. For example, some people waiting for a donor organ may need an organ with rare characteristics – which become available less often – while some people may need an organ with common characteristics.
Life expectancy boosted
Experiments with OrganITE were done using a mixture of real and synthetic data, with the real data consisting of 26 years of organ transplantation surgery in the UK (approximately 19,000 patients). Following the recommendations of OrganITE would have resulted in higher life expectancy and fewer deaths, both before and after transplant surgery, compared with other organ-allocation policies. The OrganITE recommendations raised the average life expectancy of patients significantly, by approximately 6 to 18 months (depending on which existing organ-allocation policy it is being compared with).
Predicting unbiased individual treatment effects in an organ setting is non-trivial for two reasons. First, because organs are never assigned to recipients at random, informed assignment introduces bias in available datasets. Second, as every organ and patient pairing is unique, every prediction is necessarily an out-of-sample prediction.
OrganITE solves for this by first transforming organs to categorical variables by means of unsupervised clustering. It then represents the data in such a way that it becomes much harder to predict who received which organ – effectively balancing the dataset – and handling past assignment bias. Using this balanced representation, OrganITE can make unbiased predictions of life-expectancy for every patient-organ pairing.
“It is challenging to make unbiased predictions in a medical setting, especially when a clinical trial is infeasible, as is the case in organ allocation,” says the paper’s first author, Jeroen Berrevoets of the van der Schaar Laboratory. “In such a scenario, methods like OrganITE can help tremendously by providing clinicians with balanced predictions for previously unconsidered organ-patient pairings, potentially increasing the reach of transplant surgery.”
OrganITE might be considered controversial, because it does not necessarily recommend that an organ goes to the most suitable patient (i.e. the person with the largest survival benefit from that particular organ), but instead considers the wider context and benefit to the patient population. Consider the following scenario: Patient A and Patient B both need a liver transplant. Patient A needs a common type of donor liver and is predicted to die in 4 months if they don’t receive a transplant. Patient B requires a rarer type of donor liver and is predicted to die in 2 months if they don’t receive a transplant. A liver becomes available that is a somewhat closer match to Patient A than to Patient B. Because of this, the donor liver is predicted to extend Patient A’s life by six months, or Patient B’s life by four months.
Nevertheless, given that Patient B is closer to death, will get significant benefit from the organ, and it is unlikely that a more suitable organ will become available, it may be a better overall strategy in this case to give Patient B the currently available organ. That’s because Patient A still has four months to live, and another donor liver is likely to become available that is a suitable match, because Patient A requires a common type of donor liver. See figure 1 for an expansion on this scenario (Patient A is uppermost, Patient B bottommost).
This advance in the technology around organ allocation, and its aim of maximising life years for the entire patient population, may fuel ongoing discussions around the equity of how donor organs are, or should be, allocated.
“At a societal level, decision-support tools driven by the newest ML technologies will not only maximise recipients’ population life years, but will also reduce the need for a second transplant, allowing more donor organs to be used for a first transplant,” says Dr Gimson. “Such tools will increasingly be helping to guide the patients and clinicians making these difficult and taxing decisions.”
Reference: J. Berrevoets, J. Jordon, I. Bica, A. Gimson, M. van der Schaar. “OrganITE: Optimal transplant donor organ offering using an individual treatment effect” Neural Information Processing Systems (NeurIPS), 2020