Objective
We will design two AI-enabled decision support systems that will provide guidance and advice to a physician and patient (1) prior to an islet transplant around the possible benefit of an islet transplant based on patient-specific factors as well as other factors related to the transplant therapy, and (2) after an islet transplant regarding how to predict ongoing success of the islet transplant in the months and years after islet transplantation. We will use data science approaches and several large islet transplant dataset registries, including the Collaborative Islet Transplant Registry (CITR), the University of Alberta Islet Transplant Registry, and the University of Edinburgh Scottish Islet Transplant registry. We will train new AI forecasting models on these datasets. These models will predict the benefit of the islet transplant using a new Islet Transplant Benefit Index that we will create. These algorithms will be explainable in that they will identify which factors are most relevant for the prediction of islet transplant benefit. This explainable algorithm will be useful for physicians as they explain the reasons why an islet transplant is likely to be successful or not and what the patient may be able to do to optimize the likelihood of islet transplant success. In Aim 1 of this project, we will develop a pre-transplant Bayesian Islet Graft Decision Support System (pre-BIG-DSS) framework to inform optimal therapeutic outcomes for people considering cell-based therapies to treat T1D. We will utilize the datasets to design a pre-transplant mixed effects random forest (pre-MERF) to forecast expected benefit of islet transplant at the time that an islet graft transplant occurs whereby benefit is calculated at various time points after the initial graft transplant (e.g. 1, 2, 3… n years after transplant). Using the MERF and the islet datasets, we will generate a number of digital twins (metabolic replicas of a real-world person with T1D) that will be used to simulate the benefit of various interventions undertaken in islet transplant care. We will use a Bayesian approach to identify the optimal pre-transplant recommendations that should be provided to a patient and physician to yield an outcome that is desirable for the patient. We anticipate that the pre-BIG-DSS will be successful at identifying therapeutic options that improve benefit for patients by 20-30% and that predictions of outcomes will be accurate within ±10%. In Aim 2, we will develop a post-islet transplant Bayesian Islet Graft DSS (post-BIG-DSS) that can provide ongoing recommendations about therapy following islet graft transplants to improve transplant success. We will use the ITBI to define the expected benefit of the post-transplant therapies. We will use an AI algorithm to design a set of digital twins that will appropriately respond to different changes in post-transplant therapy and accurately forecast the ITBI and other corresponding benefit outcome measures comprising the ITBI. The digital twins will then be used to design a post-transplant BIG-DSS to provide optimal recommendations for improving the outcome of the graft transplant during follow-up visits in the first several months and then annually post-transplant. In Aim 3, we will evaluate how CGM data pre- and post-transplant can help augment prediction of islet transplant therapy outcomes. We will evaluate how features extracted from CGM data from the Alberta and Edinburgh registries pre- and post-transplant contribute to prediction accuracy. We will explore traditional features from CGM and also features from a CGM foundation models (similar to Open AI’s ChatGPT) to compare impacts on performance. The outcome from Aim 3 will be to quantify how CGM data can be used to forecast pre- and post-islet transplant health outcomes and as an early marker of changes in graft function.
Background Rationale
Islet transplants have shown tremendous potential in that people with T1D who receive them (typically people with uncontrolled hypo or hyperglycemia) may achieve insulin independence and have substantially improved glucose outcomes for years following transplantation. While transplantation of islets had been an elusive dream for many years, recent developments are now making it a possibility. Over the past 20 years, protocols have been improved with different induction and immunosuppression regimens including T cell depleting agents and anti-TNF alpha agents such that approximately half of people receiving transplanted donor islets do not require exogenous insulin 5 years post transplant. Various studies have shown that islet transplantation using donor islets is superior to intensive insulin treatment whereby HbA1c can be reduced to 6.6% in the islet group vs. 7.5% in people on intensive insulin therapy with evidence of diminished microvascular complications including retinopathy neuropathy and nephropathy in the group receiving islet transplants. While islet transplants offer significant promise for enabling people with T1D to live their life without the need for using exogenous insulin delivery systems and CGMs attached to their bodies, islet transplantation does not come without problems. Poor islet engraftment and progressive decline in graft function requires most transplant recipients to resume exogenous insulin therapy, with around 30% unable to sustain graft function. Multiple factors may be impacting the success or failure of islet transplants including patient (e.g. HbA1c, total daily insulin requirement, insulin dosing therapy, duration of diabetes, and comorbidities) and donor factors as well as variations in islet preparations. People receiving islet transplants require induction agents, commonly T cell depleting agents, at the time of transplant with maintenance immunosuppressant drugs to prevent the rejection of the islet grafts. Immunosuppression is associated with an increased risk of infection and cancer, particularly skin cancer, and renal function decline. A variety of induction agents and immunosuppressant drugs are used to help prevent the rejection of transplanted islets while mitigating the risk of infection and other negative health outcomes. Other factors may directly or indirectly predict success of an islet transplant including numbers of islets transplanted, age of recipient, induction and immunosuppression agents used and their efficacy to control auto and alloimmunity. The objective of this grant is to use explainable artificial intelligence (AI) approaches to identify which patient-specific, and therapy-specific factors are predictive of positive islet transplant outcomes. We will build nonlinear mixed effects forecasting models that predict the outcome of islet transplant therapy. We will then integrate these forecasting models into a decision support framework for providing guidance to physicians and patients on how to optimize islet transplant therapy for optimal health outcomes.
Currently there are no tools available for physicians and patients considering islet transplant therapy. The decision support systems proposed will enable clinicians and patients to make informed choices before and after islet transplant. By enhancing the precision of clinical decisions, this project has the potential to maximize the success rates and overall benefits of islet transplants for people with T1D, addressing long-standing challenges in pre- and post-transplant care. The project will harness data from existing large islet transplant registries to design and validate these decision support systems. These large repositories have not yet been used to improve care for people undergoing islet transplantation. Frequentist approaches to identify the causes of islet transplant success and failure have not been successful, likely because of the large numbers of confounding factors and the relatively small number of islet recipients. The work proposed will leverage latest tools in explainable AI and the growing size of islet transplant data registries to make personalized decision support on islet transplant therapy a reality.
Description of Project
Islet transplant therapy represents an opportunity to significantly improve health outcomes in people living with type 1 diabetes (T1D) by enabling recipients to generate their own insulin without the need for external glucose sensors or insulin pumps or needles. However, islet transplant therapy is only successful in a fraction of people who receive the transplant. Currently, all people receiving islet transplants must receive immunosuppressant medication to prevent their bodies from rejecting the islets. These immunosuppressants can heighten the risk of infections and can also cause significant increases in cancer risk. Furthermore, many people receiving islet transplant therapy must receive additional transplants in the subsequent years after their first transplant because the initial transplant is no longer effective. While some people are still successfully producing their own insulin 5 years after their first transplant, a large percent of people must return to taking insulin through needles or pumps in the years after their islet transplant. In this project, we will leverage several large data sets from islet transplant registries to design new artificial intelligence (AI) algorithms that can predict health outcomes in the years after an islet transplant has occurred. We will develop a new Islet Transplant Benefit Index that is a measure of the health benefits conferred by an islet transplant. The AI models will be able to predict the likelihood of positive benefit of an islet transplant in a person living with T1D. Our proposed methodology also provides mechanisms to identify which factors relating to the patient, the therapy, and medications will most influence the predicted benefit of an islet transplant. We will integrate these AI models into a decision support framework that will provide guidance to patients and physicians prior to an islet transplant and also during the years after the transplant has been done. We expect that these AI-enabled decision support systems will be helpful for physicians and patients as they weigh the risks and benefits of undergoing islet transplant therapy. We anticipate that these decision support systems will be helpful for patients and physicians in the years after islet transplant to determine what interventions, and medication factors will be most beneficial to ensure ongoing positive health outcomes following a transplant. Finally, in this project, we will explore how CGM data from islet transplant registries collected at the University of Alberta and the University of Edinburgh can be used to improve the accuracy of forecasting islet graft survival and overal benefit of islet transplants.
Anticipated Outcome
The work proposed in this grant will result in delivery of two new decision support technologies for pre- and post-islet transplant along with a new islet transplant benefit index that can be used to quantify the risks and benefits of islet transplantation. There are currently no decision support tools available to help clinicians and patients receiving an islet transplant consider the risk and benefit of islet transplant. This grant will deliver a first-of-its-kind data-driven tool to provide decision support on recommendations for who should receive a transplant and why. We will create a pre-transplant Bayesian Islet Graft Decision Support System (pre-BIG-DSS) that will include an explainable AI-based forecasting model that will predict the likelihood of benefit of islet transplantation. The pre-BIG-DSS model will predict which patients should receive an islet transplant and which ones should not. A set of digital twins will be designed and used to forecast how use of the pre-BIG-DSS prior to islet transplantation could improve the overall health outcomes for people considering islet transplant therapy. An important outcome of this work will be the design of a new Islet Transplant Benefit Index (ITBI) that will be designed based on patient and physician feedback. This benefit index will be useful beyond the work in this grant for clinicians, researchers, companies, and patients who are interested in therapies that balance the benefits and risks of islet transplantation. The explainable aspect of the pre-BIG-DSS will yield a ranked list of the patient-specific donor-specific, and therapy-specific factors that are most relevant for forecasting which patients would most benefit from islet transplant therapy. This ranked list of factors will be important for enabling physicians to discuss the patient-specific risks and benefits associated with islet transplant therapy. In addition to the pre-BIG-DSS, the grant will yield another original data-driven tool to provide decision support on what the optimal interventional therapies, and drugs are to optimize health outcomes following an islet graft transplant. We will deliver a post-transplant Bayesian Islet Graft Decision Support System (post-BIG-DSS) that will be used by patients and physicians in the months and years following an islet transplant to choose the best ongoing drugs and therapy interventions to optimize success of the transplant. The post-BIG-DSS will be designed using explainable AI forecasting approaches that will forecast the benefit of various post-implant interventions that can optimize the success of the transplant. The post-BIG-DSS will yield a ranked list of the factors that most impact islet transplant success and benefit to the patient, thereby providing motivation for the patient and physician to utilize certain interventions after the islet transplant to optimize benefit and success to the patient. Finally, this grant will yield a new forecasting model that explores how CGM data may be used to augment benefit forecasting in islet transplant therapy and also identify islet graft health overall following a transplant. We will leverage CGM data collected from our data repositories and the latest techniques in transformer-based modeling using large CGM data sets to identify features from the CGM that are most predictive of islet graft health following implant. This transformer model will be made available to the research community for free following completion of this project.
Relevance to T1D
Continuous glucose monitoring (CGM) and automated insulin delivery (AID) systems have clearly shown some benefit to people with T1D. In particular, automated insulin is particularly beneficial during the overnight period when meals are not being consumed and when people are not exercising. Meals can be challenging for AID systems because the peak action of subcutaneously delivered insulin is typically about 60-90 minutes while the peak carbohydrate in plasma is about 30 minutes. Because of this mismatch in dynamics, the insulin is delayed in acting on the rise in glucose after a meal, and postprandial hyperglycemia results. Exercise can also be problematic in glucose management for people living with type 1 diabetes as certain types of exercise (e.g. aerobic) as well as longer duration and higher intensity exercise can cause sharp drops in glucose and dangerous hypoglycemia. The impact of AID systems on glucose control as measured by HbA1c is therefore modest, with reductions of -0.3 to -0.7% typically. AID systems have other challenges that are preventing adoption and effective use. The systems are bulky, and they require the patient to wear both a CGM and an insulin pump, both of which are injected subcutaneously. And even for those patients who choose to use them, glucose control in most people still does not meet guidelines recommended by endocrinologists.
Cell-based therapies, including islet transplantation, represent a revolutionary approach to helping people with T1D manage their glucose using donor islets without the need for using wearable CGM or insulin pump technologies. However, although islet transplant therapy has shown tremendous benefit in many people who receive it, many islet transplants fail after several years. Furthermore, there are risks associated with the therapy including the fact that immunosuppressant therapy is required for those people receiving islet transplants. For this reason, islet transplant therapy has been reserved primarily to those people with T1D who have uncontrolled hypo- or hyperglycemia. There are currently no decision support tools available to help people with T1D and their physicians understand the risks and benefit as they consider use of islet transplantation. There are no personalized tools that can help a person considering islet transplant to understand what factors should be considered when making a decision about islet transplant therapy. There are also no decision support tools to help a person with T1D who has received islet transplantation to make the correct ongoing therapy choices to ensure ongoing success of the islet transplant. The work proposed herein will make decision support tools available to people with T1D and their care providers as they make decisions about islet transplant therapy. We anticipate that this will lead to improved outcomes for those people with T1D receiving care as well as improved outcomes for those people with T1D who may choose not to receive islet transplant therapy because of the risks involved.
Artificial intelligence (AI) is exploding in many fields. We propose to leverage the work done in the field of explainable AI and Bayesian data modeling to deliver new decision support tools that will enable personalized care for people with T1D who are making choices about islet transplantation therapy both before and after islet cell transplantation. We anticipate that these tools will improve health outcomes in people living with T1D who are considering islet transplant therapy.