Scientist Stories: Mihaela van der Schaar, AI for Medicine

Hello, I am Michela van der Sar, and I would like to welcome you all to the inauguration of the Cambridge Centre for AI in Medicine. Thank you so much for joining us today. The first thing I want to do is to share with you the sense of anticipation and opportunity that we have created so quickly with the newly formed Cambridge Centre for AI in Medicine here at the University of Cambridge. In just 8 months, with support from our industry partners AstraZeneca and GSK, we have moved from a concept idea to this inaugural event. And we've done that even in the face of a global pandemic. That this rapid development was even possible demonstrates to me how excited everyone is about the enormous potential of machine learning to transform medicine and healthcare for the better. Over the next few hours, I hope you'll enjoy the wide variety of presentations we've created to tell you about our multidisciplinary center and what we plan to deliver. You'll see presentations from world-leading scientists, healthcare industry figures, and NHS leaders, and also a student research showcase. You'll notice that this webinar has a Q&A option, so if you have questions during the presentations, please type them in, in the Q&A box. We will do our best to answer them for you either during our Q&A session at the end of part 1 of the webinar or our roundtable session at the very end. But for now, please sit back and enjoy the inaugural event of the Cambridge Center for AI in Medicine. The mission statement of the Cambridge Center for AI in Medicine is to develop pioneering artificial intelligence and machine learning technologies to transform biomedical science, medicine, and healthcare. We believe that the unique sweet spot of our center resides in the interdisciplinary cutting-edge research we plan to do. We believe that medicine drives innovations in machine learning and artificial intelligence, and that in turn, cutting-edge machine learning and artificial intelligence technologies will drive a revolution in medicine and healthcare. We also aim to achieve real-world impact through collaborations with our industrial partners as well as the NHS. We will develop cutting-edge machine learning technologies aimed at enabling biomedical discoveries, enhanced omics analytics, delivering precision medicine, and enabling Next Generation Clinical Trials. Let me introduce briefly our team. I am delighted we put together an outstanding core team. Andres, who will be giving a talk after me, is my co-director, and he will introduce himself then. What is fantastic about the center is how we have managed to bring together such a diverse group of top researchers in their fields to lead on specific areas. Stefan on health management, Sarah on cellular genetics, Ewan on renal medicine, Jose Miguel in machine learning, Pietro on computational biology, Angela in health data science, Ari in anesthesia and intensive care, Alex in transplantation. Also, I would like to provide many, many thanks to Patrick, Nigel, Colm, and John, without whom this center would simply not have happened. So how is the Cambridge Center for AI in Medicine different than other endeavors in this area? First is our unique way of working. We aim to break silos doing truly interdisciplinary research. All the members of the core team are eager to learn from and teach each other. We aim to understand, formalize, and solve fundamental problems in medicine and healthcare. And as I mentioned before, in turn, we believe that medicine will drive cutting-edge research in machine learning and artificial intelligence. This is an exciting, pathbreaking, fast-moving areas with high impact. Also, another unique asset of our center is that we are going to have a critical mass of brilliant PhD students. The students are going to have multiple advisors and are going to be multilingual across different scientific areas. We will build a a multi-talented cohort of PhD students with diverse skills. These are going to be thought leaders in the area of artificial intelligence for medicine. We search for the best international students that are coming from all over the world. Why do we believe that we can do all this? First is our track record. We have an established track record in a variety of scientific areas. Here is just a small sample of our research in the last 2 years. And you can see our contributions in medicine and science. We also have done a lot of work in recent years in machine learning for healthcare. Here, we have decided to highlight our work at NeurIPS 2020, which is the most prestigious conference in machine learning. You can see a variety of our works from the core team members, and you will see our focus on machine learning for healthcare. Also, early in this pandemic, in March 2020, Even at the very beginning of the discussions about our center, the core team has joined forces to write a position piece discussing how artificial intelligence and machine learning can help respond to this pandemic. We have outlined opportunities where these technologies can help, inclusively managing limited resources, Developing personalized and effective treatment courses for each patient. Informing policies and improving collaboration. Enabling improved clinical trials, both for COVID-19 medication and vaccines, as well as for other medications for which the trials have stopped, such as cancer. And more generally, more effectively Managing uncertainty. This work was written in March 2020, and since then has also appeared in the Machine Learning Journal. Please take a look. We have proceeded since then and developed machine learning technologies for all the areas we have outlined. Also, we are different than other centers because we will focus on research impact through collaborations with our industrial partners. They are going to help us identify relevant research questions with high impact. Also, they are going to enable us to translate our cutting-edge research in practice. We aim to work closely with our pharma partners as well as NHS to enable all this. We will have an iterative process of formalizing and solving problems and assessing their performance in practice. We hope in this way to enable our partners to catalyze a revolution in medicine. So what do I mean by catalyzing a revolution in healthcare using AI and ML? Learning from a variety of healthcare data, including electronic health records, artificial intelligence and machine learning can derive actionable intelligence to inform clinical practice, to derive data-induced hypotheses, thereby catalyzing new discoveries, provide data-induced genetic associations and causal relationships, thereby providing enhanced omics analytics, and unraveling data-induced causal discoveries empowering new therapeutics. So what is really possible? What I'm really talking about? In the next few minutes, I'm going to show to you a few possible examples. So hold on tight because we will be moving very fast But I really hope you will enjoy the ride. So our vision for revolutionizing healthcare will include the following 5 pillars. We will move beyond precision medicine into what I call bespoke medicine. I will define that in just a little bit. We will empower healthcare professionals, enable improved healthcare systems, pathways and processes, enhance population health and public health policies, and catalyze new discoveries, both in terms of understanding disease and disease pathways, as well as improved therapeutics. Let me start with a quote from William Osler, the father of modern medicine. If it were not for the great variability between individuals, medicine might as well be a science, not an art. We plan to use machine learning and artificial intelligence to turn medicine from an art into a science. Medicine is hard because patients are complex. They have different genetic background, environmental exposures, lifestyles, histories, and interventions. This translates into different risks, different risks over time, variation in symptoms, different health and disease trajectories, and responses to treatment. Personalized medicine has aimed to address these challenges by fitting the patient to a pattern. In bespoke medicine, we go one step further by recognizing that we need to adapt to changes in patterns with age, lifestyle, onset of new conditions, and progress in course of treatment. So we will have many patterns that the patient is going to have over the course of his or her lifetime. How are we going to find these different patterns and match patients to patterns? Using machine learning. So bespoke medicine is going to enable us to provide a comprehensive patient view. A long-term view that depends on the context, but also a broad view because it will consider the various competing risks of a particular patient. And it will enable us to look longitudinally to a particular patient, consider the past, present, and future of a patient, thereby providing a lifetime of care. We are going to develop using machine learning risk scores, competing risks, personalized screening and monitoring methods, advanced diagnostic support, longitudinal disease trajectories that is personalized, and individualized treatment effects. Thereby, we are going to be able to move to bespoke medicine. Also, we are going to be mindful that while different diseases may seem different, we can use lessons learned from one disease and apply them across multiple diseases. Clinical problems which bespoke medicine can address include, but are not limited to, data-driven discovery of the interplay between multiple factors that affect the health and disease of a patient over time. Better quantification of disease by enabling data-driven disease characterization using machine learning, identifying disease states and durations in the different states, thereby enabling early diagnosis. Unraveling causal pathways— inclusively evolution and co-evolution of symptoms and morbidities. Dynamic temporal phenotyping and forecasting that enable us to identify the evolution over time of different events of interest for a class or a subclass of patients. This will also enable us not only to understand better disease, but to make accurate forecasts for a particular patient. We are also going to use artificial intelligence and machine learning to estimate individualized treatment effects and do counterfactual scenario planning, thereby identifying individual variation in efficacy, side effects, and interactions. We know that patients are complex with multiple risks, so current approaches which assess one risk at a time independently are not enough. We aim to develop a holistic view of the patient health, which includes many risk scores for many clinical conditions. But the challenge in order to do that is that there is no one-size-fits-all solution for this in terms of machine learning. So the question is, which machine learning model to choose? And for what type of condition. We also need to issue not only accurate predictions, but also interpretations and explanations associated with these accurate predictions, such as the clinicians, as well as the medical researchers using these machine learning tools can most benefit from them. Moreover, we need to issue trustworthy predictions, such that we know when we can trust the predictions made by a machine learning model and when we cannot. When can we transfer, for instance, the knowledge that we have acquired in a particular population, for instance, here in the UK, and when can we transfer it accurately to the United States, where the population may be different, may have different characteristics, and also standards of care? Last but not least, we are interested in providing reproducible, machine learning methods such that you can take the technologies that we have developed here and rely on their performance when applied in a different ecosystem. So what can be achieved? A few years back, we have designed AutoPrognosis, which is a tool for crafting machine learning-enabled clinical scores. AutoPrognosis takes clinical data and builds entire pipelines to issue clinical predictions. These pipelines contain multiple stages: data imputation, feature processing and dimensionality reduction, classification, and calibration. AutoPrognosis selects at every stage of the pipeline the best-performing algorithm for this particular dataset, as well as optimizes the parameters and hyperparameters of the different algorithms. Autoprognosis is not only issuing predictions, but also explanations associated with the various predictions. In this way, the clinicians and medical researchers using autoprognosis can understand, debug the different predictions made by autoprognosis. Autoprognosis is also able to learn the value of information for a particular clinical prediction. Auto-prognosis is in this way able to identify what type of tests and information needs to be collected for a patient or a class of patients in order to achieve an accurate prediction. It may be that for a subset of patients, for instance, the diabetic patients, a different set of measurements and observations are necessary in order to issue an accurate prediction. AutoPrognosis is able to learn what data needs to be collected in order to issue such accurate predictions for subsets of patients. In addition, AutoPrognosis is also able to learn the value of modeling. We are able to understand for a particular clinical data whether regression models are going to be sufficient for an accurate prediction, or we need machine learning nonlinear models to achieve significant performance improvement. We are able to learn not only on the basis of the local clinical data, but also able to learn from other complementary cohorts. At other locations and even other countries. However, this is challenging because data coming from complementary cohorts may contain patients that may have different characteristics. They may be older. They may smoke. They may have different type of morbidities and comorbidities. And also the clinical practice may be different. So we need to learn in the presence of these differences. For this, we have developed cutting-edge meta-learning and transfer learning methodologies that are able to deal with the distributional mismatch, feature mismatch, and clinical practice mismatches in the complementary cohorts in order to integrate this information to enhance predictions locally. Auto-prognosis is able in this way, together with these various other accompanying technologies, to provide clinicians with a powerful tool on the basis of which they can build a comprehensive view of the patient at hand. And they can use this tool easily without prior machine learning expertise. We have validated autoprognosis in a variety of clinical settings: cardiovascular disease, cystic fibrosis, breast and prostate cancer, etc. And most recently, we have used autoprognosis in the fight of this pandemic, COVID-19. I'm going to talk more about that towards the end of my talk. Let us now switch gear and move from clinical scores to polygenic risk scores, where the focus is to use genomics information to predict whether a patient or a class of patients may have an increased risk for developing cancer, a particular type of cancer, or cardiovascular disease. State of the art in polygenic risk scores so far have been based on linear models. And it was a long-held belief that linear models are sufficient for polygenic risk scores, and not much improvements can be derived on the basis of nonlinear machine learning models. But the reasons behind this belief were that the polygenic risk scores were developed on datasets that were relatively limited. And the reason behind that is, for instance, the UK Biobank has only a limited set of patients for which we have phenotypic information, for which we have labeled data. Instead, what we have done, we have used a larger dataset that contains unlabeled data as well, patients for which we do not have the phenotypic information. So we are able to learn from both the labeled as well as the unlabeled data in a resource such as UK Biobank in order to craft polygenic risk scores. And what we have done is for that we have used a novel semi-supervised and self-supervised learning methodology to craft such polygenic risk scores. This represents state of the art in machine learning. Because unlike image and language data, where such methodologies have been developed before, where the spatial and sequential correlations in such data are well known, in the genomic information, we do not know these correlations. We need to learn them on the fly on the basis of the available data. In our new machine learning technology called VIME, we have enabled this methodology to learn such correlation. Then using ViME, which is a self- and semi-supervised learning technology, we have shown that we can achieve state-of-the-art performance on genome-wide polygenic risk scores, where the performance of this technology significantly outperforms that of linear models, thereby showing the power of machine learning to improve linear polygenic risk scores. Let us now move one step further, beyond predictions into personalized therapeutics. An important frontier to which the center is going to pay a lot of attention is the use of machine learning to improve Clinical trials. Clinical trials represent the golden standard for estimating average treatment effects. But unfortunately, they are time-consuming, they have enormous costs, they have relatively small sample sizes, and often enroll non-representative patients. For instance, patients with comorbidities are often not enrolled. In our work, we have used machine learning to show how clinical trials can be improved. For instance, we can use machine learning to identify which patients are most effective and need to be enrolled in a particular clinical trial, thereby determining enrollment eligibility criteria. Also, machine learning can determine dosage and what type of dosage we should provide to a particular class of patients. And also, as the trial is proceeding, identifying in subsequent stages of the clinical trial what patients are most effective, are most served by this particular therapeutic, and how we should change and identify subclasses of patients to be enrolled in subsequent parts of the trial. In recent times, together with clinical partners, we have shown how machine learning can play an important role in the era of COVID-19, both in clinical trials for COVID-19 medication, but also other treatments such as cancer. Another area where machine learning can play an important role is using observational data to estimate the effect of a treatment or intervention on an individual. Upon being diagnosed with a particular disease, we would like to identify for that particular patient what treatment would be most effective in treating this disease. We would like to identify when to treat a patient, what type of treatment we should use for a particular patient, and if the particular treatment stops being effective, when we should switch to another medication and what medication. We are able to learn all that from observational data using machine learning. So we are aiming to provide personalized therapeutics by using machine learning to learn from real-world observational data. And we have done in recent times a lot of work in this area. This particular technology is called causal effect inference. And represents a complicated machine learning problem where we aim to estimate counterfactuals. We are answering what-if questions. And this is a challenging problem because we never observe counterfactuals. So this is not a simple supervised machine learning problem because we do not have explicit labels. In the training set, we may have patients that were treated or not treated and only observe the outcome for that particular decision. We do not observe what would have happened if the patient was treated differently. We need to estimate that while not having access to an explicit label. Another challenge associated with estimating causal effects from observational data is that this data, unlike in the clinical trials, is biased by clinical practice. Doctors do not close their eyes and determine whether a patient should be treated or untreated. They are using their clinical judgment in making these decisions. So we need to learn in the presence of bias. How to solve these challenges and what are the opportunities associated with us is described for us by us in a recent tutorial where we review a lot of different technologies, inclusively many developed in our lab where we show the role that individualized treatment effects using machine learning can have when learning from real-world patient data. Today I'm not going to focus on giving an example from cancer or cardiovascular disease, but instead I'm going to give an example from the area of transplantation, which is a very challenging decision in medicine, whether we should transplant a particular patient given a particular available organ and when. This is a collaboration with Alex Kimson from our center, and more details will be given about this particular machine learning technology for transplantation by Arun Berriwitz in the student showcase. But the focus here is to use machine learning methods to estimate causal treatment effects at the individualized level for transplantation. We are using patient features and rarity of the required organ to estimate individual effects of the organ and improve transplantation policies based on the characteristics of the patient. In this way, we are moving from a prediction setting into a causal effect inference problem. A few years back, I've developed machine learning technologies for personalized survival prediction. And our work received a lot of attention because it represents state of the art in prediction for transplantation in heart transplants. Which is a very complicated area. However, we would like to move beyond predictions nowadays and estimate the effect of different types of interventions for such patients. So we have here, for instance, what is the survival probability for patients that are suffering from heart failure and are on the transplant list?. And we see their survival probability if they are given an LVAD or not. LVAD stands for left ventricular assist devices, and these are types of implanted heart pumps that represent either bridge therapy or even destination therapies for patients with heart failure. So we would like to answer the following question on an individualized basis. Would a particular patient, given their unique characteristics, benefit or not from having an LVAD? And this is an important question that cannot be solved using clinical trials, which will be unethical. And also this particular LVAD surgery costs $700,000 or more in the United States. So this can only be done on a small subset of patients. Yet we would like to identify which are the subset of patients that will most benefit from this complicated intervention? So we would like to answer this on the basis of observational data. So this is highlighted here. We have a patient which is female and is 54 years of age. And she has been for a quite large number of days on the list for transplantation, for heart transplant. And what we can see is that she has died after a couple of years while waiting for an LVAD or a transplant. And the question we would like to answer is, would have this patient benefited if she would have received an LVAD? And when should she have received an LVAD in order to have benefit? We are able to answer such counterfactual scenarios analysis and are able to answer such questions using causal effect inference enabled by machine learning. We are able to estimate that if these patients would have received in 2009 an LVAD, she would have survived 3.5 years more. It is such complicated questions that we can answer using machine learning. So in this way, we can provide a key step towards a final priority allocation scheme for transplantation. And again, this is not possible using clinical trials, which are expensive, unethical, and often rely on small sample sizes. So by developing machine learning technologies for causal effect inference, we are able to learn such individualized survival benefits from different types of interventions, thereby solving challenging causal inference problems. And this is again useful not only for transplantation, but for determining whether we should use a particular type of drug for cancer, when, and in combination with what other interventions or drugs. We are able hence to build comprehensive individualized healthcare atlases personalized to a particular patient on the basis of the observational data. We are able in this way using machine learning to forecast expected biomarker trajectories, thereby being able to identify what will happen to this patient over time and when do we need to screen them to identify a change in their health or disease trajectory. We are also able to estimate projected risk, inclusively risk from different morbidities and comorbidities, and do time-to-event analysis. We have done so for complex diseases together with Andres Floto, for instance, for cystic fibrosis. We have done this type of individualized disease atlas by building machine learning models that are clinically actionable and are able to accurately predict the patient-level trajectories. We are able to learn from complex and irregularly sampled clinical data, inclusive data that is not missing a trend but that is informatively sampled. We are able to learn from limited supervision with only limited clinical annotations and from expertise from clinicians. We are able to learn on the basis of the available historical data and identify what past events from the history of a patient matters, thereby learning holistically. Also, patients are heterogeneous, so personalization matters. We are also developing methods that are interpretable and hence actionable for clinicians. And we have done that in a suite of past works. And we are developing in the center new technologies to further empower clinicians. Most recently, together with Ari Ercole from the center, we have developed Clairvoyance, which is the first unified end-to-end pipeline for clinical decision support. Clairvoyance provides preprocessing, imputation, uncertainty estimation, calibration, and interpretations. And it does so for personalized predictions, personalized predictions over time, personalized treatment planning, and decision support, as well as personalized monitoring and screening. We have developed this suite, and we have provided it freely to everybody who wants to use it. And we hope that in this way, we can improve the work in machine learning in healthcare, also providing reproducibility, where people can bring their machine learning methods into Clairvoyance benchmark them against existing algorithms, and thereby we can improve the code base on machine learning for healthcare. We also have identified methods to use machine learning for providing personalized and dynamic predictions for morbidities and comorbidities, which represents an important frontier in healthcare. Current state-of-the-art technologies for identifying, understanding, and issuing predictions for morbid populations are usually one-size-fits-all population level and static. Networks are built on how diseases interact that are at the population level and only done at a particular moment in time. Instead, in our work, we have developed machine learning models based on deep diffusion processes, which are bespoke to the particular patient. So we identify different classes of patients and how the morbidities may change for these different classes of patients over time, thereby moving from a one-size-fits-all to a personalized approach. Also, we are able to do this dynamically and identify how the different morbidities are going to lead to new morbidities over time if they remain untreated. And we look at the duration and the order of the different morbidities and the role they play in the development of a particular new morbidity for a class of patients. Let us now move to another area where machine learning can help, that of empowering healthcare professionals. We are building in the center personalized machine learning assistants to support clinicians, to enable reflective practice as well as augment their decision-making, thereby empowering them. Different clinicians may see different types of patients with different types of characteristics, and hence these machine learning assistants need to be mindful of both the specialty of the clinician, but also the types of patients they are seeing. Also, the machine learning technologies that are used to support such clinicians need to be explainable and trustworthy, such that the clinicians know when they trust these predictions and when they cannot. Last but not least, machine learning can promote multidisciplinary clinical contributions. It can enable clinicians, for instance, dealing with complex patients such as cancer patients, to identify what information they should exchange and what type of knowledge they need to exchange in order to better treat a particular patient. So machine learning can significantly advance multidisciplinary team meetings and enable clinicians to cooperate with each other using and empowered by machine learning to provide better care. For their patients. An example is provided in the area of cancer, and we have developed here together with Cem Rashbas from NHS Digital a comprehensive decision support tool for patients with cancer. Cem is going to talk about our research in this area developed a few years back in his talk later today. In order to turn such machine learning models into actionable intelligence, We need to have risk or understanding. We need to have transparency. We need to avoid implicit bias. And we need to enable users to distill insights and new knowledge from the learned model, thereby enabling new discoveries. Also, as I mentioned before, the machine learning models need to provide also quantification of the model's prediction uncertainty, such that the clinicians know when to trust this and when they should not. One particular technology that we have developed to enable this is the use of symbolic metamodels that are capable to demystify any black box model and turn it into a white box transparent model that is interpretable and understandable. What type of black box model? There could be any deep learning model, random forest, XGBoost, et cetera. And symbolic meta-models are able to output a transparent function describing the prediction of such black box model. What is also important to note is that we do not need to know what is in the black box model. We do not need to have access to the intellectual property of this black box model. And meta-models need only query access to the trained black box model. Thereby, we are able to turn black box model into interpretable solutions that are user dependent, thereby moving beyond interpretability into explainability, which enables users to seek different types of understanding. For instance, The clinician may want to interrogate the black box model using meta-models to understand why a treatment is recommended for a patient. A researcher may want to understand why a particular data-induced hypothesis was issued by a machine learning model. And a patient may want to understand why a particular lifestyle recommendation has been made to him or her to reduce the risk from cancer. Machine learning can also help improve systems, pathways, and processes in healthcare. And our center is going to develop technologies for that. For instance, machine learning can improve healthcare pathways by enabling us to provide equal treatment across venues. Machine learning can help understand the variation in care and decrease the variability in care centered on the clinician. While delivering bespoke medicine, thereby increasing the variability in care centered on the patient. We are able to do that using cutting-edge inverse reinforcement technologies that we are currently developing. We are integrating and curating data sources and learning uses using a new diverse set of machine learning models that are able to enable us to identify how to best treat a particular patient in the most effective way at a particular location. We are also able in this way to use machine learning to improve cooperation, interaction, and learning across healthcare systems such that we can provide effective and cost-effective care. An example can be given from the area of COVID-19. Early in this pandemic, together with Ari Ercole, we have identified that there was a big variation in ICU mortality between the different centres in England. And one of the reasons behind that were that the resources were used differently across the different hospitals. So together with NHS Digital, we have built a machine learning-enabled system to help clinicians plan and manage better COVID-19 treatment resources. We aim to use machine learning to empower clinicians as well as managers to make decisions that are helpful for their patients. We use the personalized data to forecast what is the personalized risk for a patient to die from COVID-19, and what are their benefits of using different types of resources such as ventilators or other treatments, inclusively ICU beds. And this is an important question, not only at the beginning of the pandemic, but even now when we have the third wave of this pandemic. And we have used AutoPrognosis to enable that. You can see here how AutoPrognosis is issuing significantly improved predictions on the benefit of a patient to have— they may have from ventilation. You can see that the AutoPrognosis is doing better than regression methods or the Charleston Comorbidity Index. Also, AutoPrognosis is able to identify the value of information from integrating different types of features of the patients. We have built together with NHS Digital a demonstrator of this technology that is showing how resources may be needed over time in a particular hospital, such that the clinicians and managers of different hospitals can better plan their available resources. Now let me move to the use of machine learning to inform public health policy and population health. Machine learning can help us discover and disentangle public risk and risk factors. We can do causal discovery on the basis of diverse healthcare data from various populations. We can also identify using machine learning population risk assessment and determine personalized risk. Another area where machine learning can play a significant role is in data-driven guidelines and protocols and standards for healthcare. For instance, personalized screening guidelines. In recent years, we have shown the important impact machine learning can play in determining who to screen, when to screen, using what technology to screen, and how often to screen patients at risk of, for instance, cancer. Currently, patients are screened in a one-size-fits-all fashion for breast cancer. In the UK at age 50, in the United States at age 40. But the reality is that this is very ineffective. Instead, we need to move to a personalized approach to screening, where some women, given their unique risks, may need to be screened much earlier, such that we can identify their disease early, while other populations that are at lower risk may need to be screened later, and possibly using different types of technologies. And may need to be followed with subsequent screening at different intervals, thereby building more effective ways to identify cancer early. Another area where we can learn using machine learning for population, to improve population health, is cross-country learning and interventions in pandemics such as that of COVID-19. We have built on the basis of observational data from many countries a machine learning model for estimating the effect of non-pharmaceutical interventions in different countries. What type of lockdown policies we should implement and when should they be lifted and what is the effect of them. We have shown in this work the importance of learning from diverse datasets compiled across the entire variety of countries, but the impact at the local country level. I hope I have showed you in this way a variety of machine learning technologies and the role they can play in catalyzing a revolution in healthcare, whether it's enabling bespoke medicine, empowering healthcare professionals, improving systems, pathways, and processes in healthcare, or catalyzing new population health and public health policies. I hope you will stay tuned to the breakthrough research that we plan to do in the Cambridge Center for AI in Medicine. We plan to write a variety of position papers shaping this area, develop cutting-edge AI and machine learning methods and software that we plan to share with the larger community. Build demonstrators that show the power of these technologies. We will have also tutorial sessions and engagement sessions presenting this type of technologies. So I hope you will engage with us and stay tuned. Thank you very much.