Mihaela van der Schaar
Nick MaxfieldEarly detection and diagnosis (ED&D) is arguably one of healthcare’s true “holy grails.” This is something we outlined in depth in a recently-published page on our site, in which we explain the importance of ED&D and highlight a range of machine learning approaches that could catalyze transformative progress.
In our lab’s February 8 and March 10 Revolutionizing Healthcare engagement sessions for clinicians, Mihaela van der Schaar and two international panels of clinical experts explored how AI and machine learning can help transform ED&D.
This page provides an accessible and easy-to-digest summary of some of the highly insightful and informative points made by our panelists in both roundtables.
February 8 panel:
- Prof. Hari Trivedi (Assistant Professor, Dept. of Radiology and Imaging Services & Dept. of Biomedical Informatics, Emory University School of Medicine)
- Prof. Parag Mallick (Associate Professor of Radiology, Cancer Early Detection Canary Center, Stanford University)
- Prof. Tony Ng (Head of Comprehensive Cancer Centre, School of Cancer & Pharmaceutical Sciences, KCL; Professor of Molecular Oncology, UCL)
- Prof. Willie Hamilton (Professor of Primary Care Diagnostics, University of Exeter; Director of Research, CanTest Research Collaborative)
- Prof. Yoryos Lyratzopoulos (Professor of Cancer Epidemiology & Group Leader for Epidemiology of Cancer Healthcare & Outcomes Group, UCL)
March 10 panel:
- Alberto Vargas, MD (Chief Attending, Department of Radiology, Memorial Sloan Kettering Cancer Center)
- Prof. Henk van Weert (Professor, General Practice & Family Medicine, Amsterdam UMC; and Research programs in oncology and cardiovascular diseases)
- Prof. Stephen Friend (Visiting Professor of Connected Medicine, University of Oxford; Chairman and co-founder, Sage Bionetworks, 4YouandMe)
We have outlined several potential uses for ML in ED&D; which do you find useful, which are less so, and what did we miss? Would you care to share any thoughts/ideas on areas that are seldom talked about?
Prof. Tony Ng
Head of Comprehensive Cancer Centre, School of Cancer & Pharmaceutical Sciences, KCL; Professor of Molecular Oncology, UCL
From the clinical point of view, I find your dynamic prediction section particularly helpful. We cannot just make predictions based on one time point, because the disease will evolve. Take lung cancer for example: we know there are different states of the disease where people get bronchoscopies but in follow-up the tumors disappear. Predictably, there are some patients who, upon follow-up, will reverse the state. And we want to understand to some extent, as clinician scientists, what reversed that state.
This really links to some additional thoughts about causality. Some of the things we measure in blood or bronchoscopy samples might indicate something about causality and therefore link to treatment. The problem is that we have too many parameters, too many covariates, and therefore it’s hard to know which ones to treat—even though they might appear in your explainable model, which of those parameters is actually causal?
Finally, I was very interested in your statement about how you’re treating or detecting one disease at a time for any given patient. That’s certainly the state of the art int he NHS, but there are also pan-cancer clinics coming up online. These are rapid diagnostic clinics that detect cancer in patients with vague symptoms that don’t specifically pronounce to be potentially lung cancer or colorectal cancer. I certainly find your methodology very intriguing in terms of whether we can tap into the pan-cancer clinics in the future.
Prof. Parag Mallick
Associate Professor of Radiology, Cancer Early Detection Canary Center, Stanford University
When I look at the list of applications, I think one of the most exciting opportunities is taking advantage of early data. When we look at how people are diagnosed today, it’s usually based on incidental factors when they’re not feeling well (for example, presenting with a cough). If we can go back and trace a year, two years, five years earlier to look at symptoms that may have (on their own) seemed uninteresting, I think there’s tremendous value in being able to go into the pre-history to change the risk.
I think this follows along with dynamic risk prediction, but really taking into account lifestyle factors: we also have the emergence of all these monitoring gadgets that people have now, and interpreting that data is quite complex. There are really great opportunities for incorporating all of those orthogonal data elements (lifestyle factors, diet, etc.) in a level of detail that we just haven’t been able to do before.
In terms of what’s missing from the list: really, I think there’s a whole set of machine learning tools that are operating not on clinical data but, for instance, reading the scientific literature to identify correlations, gene relationships, gene biomarkers, and so forth, and extract from this unstructured data some useful and interesting information. So I suspect that machine learning tools will also play a huge role in really helping us get value from data that has been collected over the last many decades, in novel ways, to allow us to do a better job of predicting, typing, and phenotyping.
Prof. Willie Hamilton
Professor of Primary Care Diagnostics, University of Exeter; Director of Research, CanTest Research Collaborative
I’m a family practitioner, so my job is twofold: it’s identifying who should be investigated for possible cancer (and the key word is “possible”), and how to test them (i.e., whether to test them in my general practice office or whether to refer to specialists).
The question of who to test is relevant because we should be in the position of not just being able to adjust treatment according to the individual, but also being able to adjust investigation according to the individual. We do this in the U.K.’s healthcare service by putting people at low risk on long waiting lists, and people at high risk on short waiting lists. But actually there’s different tests: I could do a fecal immunochemical test for someone at low risk of colorectal cancer, while immediately sending someone for a CT scan or colonoscopy if they’re at high risk.
In terms of how to test, I can give an example of how powerful Mihaela’s type of work is. I leant my datasets to someone doing a Ph.D. to test run fairly simplistic machine learning on them. They managed to push the ROC performance up from the high 80s to the mid 90s. From a patient perspective that is magnificent. The patient and health economic benefit is enormous.
Prof. Hari Trivedi
Assistant Professor, Dept. of Radiology and Imaging Services & Dept. of Biomedical Informatics, Emory University School of Medicine
The thing that jumps out to me the most is personalized screening, which I think would be of extreme value. I work in mammography, and we’ve been working on developing a dataset of patients where I practice. Our patient population is 50% African American here. That patient population is not represented in the vast majority of any of the trials that determine screening for breast cancer (the efficacy, risk, and decision to do annual screening, etc.).
If you look at the existing models that are used (the Tyrer-Cuzick score in particular, but also the Gail model), they’ve been amended several times but still have a gap in performance across races. The issue is twofold: one is data availability for these patients, and the other is understanding that a one-size-fits-all screening paradigm across patients doesn’t make sense when we have more powerful techniques. It’s hard to make risk predictions when you don’t even have that data because a patient doesn’t even come in.
Scale-wise, one of the biggest impacts we could have will be if we can get the right patients in at the right time while conserving resources, and then tailoring screening schedules for these patients on an individual basis. Bringing these patients in at the right time could allow us to catch thousands and thousands of potential extra cancers.
Prof. Yoryos Lyratzopoulos
Professor of Cancer Epidemiology & Group Leader for Epidemiology of Cancer Healthcare & Outcomes Group, UCL
Mihaela, your work highlights the need for good electronic health record (EHR) data. By this, I don’t just mean data that are well-curated and accessible or available; but rather being able to record the fullness of symptoms in a way that is efficient to the consultation. Similarly, the EHR and system interface, the environment in which the consultation happens, needs to enable a full history (ideally a coded history) to be taken.
There are many EHR systems around the world that are essentially deserts; they are isolated entities that don’t speak to each other, so I expect there are some great gains to be had through AI contributions to interoperability, and making very badly architectured systems distil more evidence for the work that you and other colleagues here are doing.
Prof. Stephen Friend
Visiting Professor of Connected Medicine, University of Oxford Chairman and co-founder, Sage Bionetworks, 4YouandMe
As clinicians, there is a strong tendency to focus on “what,” not “why.” When people do clinical studies, they get very excited about sharing what they studied in a particular population, and what happened within that population. And others can pick up off such research.
This approach is great for generating papers and funding; the difficulty is that we are functioning as “medical alchemists.” After thousands of years, chemists realized that they needed a periodic table to tie causality to what’s actually happening.
There is a serious opportunity to spend more time posing the questions that we’re working on in terms of “whys” in order to get to the rules. It is extremely hard to look at whether a particular finding is generalizable if you don’t know the context and what is driving the thing you are studying.
We will be missing an opportunity until we take the time to ask whether there are rules underlying what is going on. I hope it’s obvious how machine learning could help in the identification of those rules. Maybe this could even be done in such a way that, when you start to look at enough individual particulars, you can begin to have a “contour map” of what is similar (in causal terms) that might allow you to find those rules.
Another area that is seldom talked about is data protection. I think there’s an illusion of security when people claim that “we will protect your data.” Through working in industry and academia, I know that people who are working with machine learning using EMR data, social media data, or other types of data have a way of understanding (beyond data and knowledge) why people do what they do. This is also true of wearables, where we are given enormous amounts of in-depth, semi-continuous data. Some of us need to be working on understanding the inadvertent consequences of working with insights into behaviors and how these insights could be used in ways that might not be what we meant.
Prof. Henk van Weert
Professor, General Practice & Family Medicine, Amsterdam UMC, and research programs in oncology and cardiovascular diseases
General practitioners and most doctors are not looking for things that might happen, but for things that have already happened. I think that’s a very important difference, as it means we can predict what will happen from patient history. We don’t have to look forward, because after all a patient already has cancer by the time an early diagnosis is made. That makes it easier to target your searches: we’re not talking about populations here, but individuals within a population. Specifically, we’re looking for patients with high risks of diagnosis, or rather patients who have a diagnosis without knowing it yet.
Having said that, any one type of cancer, in an epidemiological sense, is a rare thing to have happen in a patient. And for every prediction you make, you’ll have more false positives than true positives. This is something you have to explain very well to people working on this, as well as patients. If we consider everything that has already been discovered with regard to early detection, we see that GPs use the instruments at their disposition very well. But at the moment they refer a patient for cancer, that patient (at least in the Netherlands) will have about a 50% chance of having a disseminated cancer at the time of referral. Bear in mind that when a patient is referred by a GP, this will be done within about 2 weeks after presentation with an indicative symptom.
So, rather than just looking at indicative symptoms, we need to be able to look for other details further back (i.e., a year or more) in time in a patient’s history and look for cues that can identify patients at risk earlier. The cues should not be individual symptoms anymore. However, we are looking for things we don’t know yet—and we don’t know what we’re looking for! The good thing, I think, is that machine learning is made to look for things we don’t know yet.
Alberto Vargas, MD
Chief Attending, Department of Radiology, Memorial Sloan Kettering Cancer Center
I agree that machine learning is made to look for things that we don’t know and can’t see. This is particularly relevant to my specialty as a clinical radiologist in a cancer center. Radiology has exponentially grown over the last 20 years, with a fourfold increase in the number of exams that are done. This means that almost every single patient going through a hospital will have some form of imaging.
Imaging is very effective at rapidly providing useful, actionable information for clinical indications, with the evaluation being targeted to answering a clinically posed question (such as a CT to identify pneumonia). This contributes massively to patient workflow. But what we don’t look at is the hundreds and sometimes thousands of images included in one individual exam. We couldn’t do this, even if we wanted to—after all, fully scrutinizing one image per exam is challenging enough! There is an immense amount of data going unanalyzed, unused, and unrepresented. I feel that this is where the major opportunity lies, with regard to images. We can try to grasp information that may be there but is not visible to the naked eye or assessable for a human radiologist who needs to get through a certain number of exams in a working day.
Additionally, the era of expansion in radiology is over. We will continue to grow to some extent, but this rate of growth will not continue. We have entered an era of integration. The imaging information is useful and should be scrutinized to a greater extent, but one of the key sources of value for machine learning will be trying to ensure that every factor is considered when each individual source of data is being evaluated.
What types of data do you think will be needed to inform machine learning methods for ED&D (whether it’s wearables, other new types of devices, or something else)?
Prof. Tony Ng
Head of Comprehensive Cancer Centre, School of Cancer & Pharmaceutical Sciences, KCL; Professor of Molecular Oncology, UCL
The kind of data we’re looking at for the rapid diagnostic clinic are symptoms, which are actually very important, as is using NLP to actually mine those data.
Apart from that, the next bit of data is imaging, particularly for those patients who don’t get a diagnosis straight away. They will have follow-up imaging such as lung CTs and so forth, so capturing longitudinal images is very, very important.
Finally, we are in a time when in the U.K. GRAIL is doing pan-cancer tests and many companies are coming to the U.K. because of the NHS; all those blood test parameters, if you can get hold of them, would be very important.
Prof. Parag Mallick
Associate Professor of Radiology, Cancer Early Detection Canary Center, Stanford University
My thought is less about the actual data modality itself, and more about the quality and annotations on the data. Machine learning tools depend on having exceptional labels. For image data, knowing what to look at is very important (so, for example, determining where is the cancer in this pathology section?).
To me, that would be the focus; not just on volume or type of data, but really on quality and annotation.
Prof. Henk van Weert
Professor, General Practice & Family Medicine, Amsterdam UMC, and research programs in oncology and cardiovascular diseases
The short answer is that we don’t know what types of information may prove useful!
Every new piece of information that becomes available can help in one way or another, but we don’t know beforehand. You have to explore whether or not it contributes to making a diagnosis or choosing a therapy. Afterwards, you can tell whether it helped or not. That’s the soul of machine learning, to some extent: that you don’t know beforehand.
Alberto Vargas, MD
Chief Attending, Department of Radiology, Memorial Sloan Kettering Cancer Center
Specifically with imaging, integrating the imaging with patient-reported outcomes is a very important untapped field, in my opinion.
Conventionally, a patient would come in for an outpatient visit every 3 or 6 months, and we would ask them how they’ve been. Now, though, there’s increasing use of self-reported patient outcomes that a patient can input on a daily or weekly basis that can more accurately monitor their state. Plugging that into the data you can generate from imaging and other sources is probably one of the key points to consider in that respect.
Prof. Stephen Friend
Visiting Professor of Connected Medicine, University of Oxford Chairman and co-founder, Sage Bionetworks, 4YouandMe
When you get to n=1 predictions as an alternative to making aggregates, you will need much deeper data than we currently have.
We’ve been interested in wearables as a way to have thousands of dimensions for data that are being collected many times a day in order to get some density and depth. Many people are focusing now on wearables as a source of physiological signs, and they’re assuming that patient-reported outcomes will tag along. But what happens if you use multi-modal approaches to go from sensor data to physiological data and symptoms?
It would be very nice to have an objective way to track physiological symptoms, because then the interventions can come out. This way, clinicians can take the ambiguous self-reported “I feel” symptoms and tie them to something real. With wearables, I don’t think the value of information lies in individual physiological signs. I don’t care about heart rate or activity alone. I want to know how we weave these together: how can you take between 100 and 1,000 dimensions of fatigue, for example, and weave them together to provide an objective assessment? I think this is an important role for machine learning to play.
We have previously shared a framework for machine learning for healthcare. In which of the 4 areas of this framework do you think machine learning would add the most value for ED&D?
Prof. Yoryos Lyratzopoulos
Professor of Cancer Epidemiology & Group Leader for Epidemiology of Cancer Healthcare & Outcomes Group, UCL
I like all of these areas very much (as, I’m sure, does everyone else here).
I would say that the area that is perhaps under-catered and under-served is systems, pathways, and processes in the bottom-right. I think an AI-informed contribution to systems, pathways, and processes will be terrific.
This can involve a lot of things, such as applying human factors engineering to EHR interfaces, so that treating teams and patients themselves can generate high-quality, plentiful data. A lot of the information currently is not really reflected and phenotyped in the system.
Additionally, I would add to this area a second dimension: the stewardship and curation of the EHR system itself. It has been terrific to see recent AI applications where the code becomes runnable in very different or foreign EHR infrastructures.
Prof. Hari Trivedi
Assistant Professor, Dept. of Radiology and Imaging Services & Dept. of Biomedical Informatics, Emory University School of Medicine
I think each of the corners are extremely valuable, I really feel strongly about risk prediction, mostly because of the scale of impact.
When you think about the challenges, something I wonder about a lot is: at what point do you just end up with individual genotypes and phenotypes per patient? How far can we get before every patient just becomes a unique n of one? At that point, is the challenge we face a technical one, or do we lack the labs and assays and genetic tests and understanding to even get beyond that? In essence, how much further down the road do you have to get in order to have an accurate prediction on one extra patient?
We recently wrote a paper on predicting hospital readmissions, where we predicted a patient’s return to the hospital within 30 days from discharge, based on insurance reimbursement data. You can get up to 70% or 80% because you can easily group patients into large cohorts, but after a year of developing additional models of this space, and concluded that either we lacked data to discriminate against these additional patients, or that perhaps it is simply not possible for us to get to a point where each patient is individually unique.
Since these challenges are not yet answered (i.e., we don’t know if we have insufficient data or whether the barrier is a technical one), I think for the time being it’s risk prediction that has the potential for impact at a much bigger scale.
Prof. Willie Hamilton
Professor of Primary Care Diagnostics, University of Exeter; Director of Research, CanTest Research Collaborative
I like all four of these areas but I think the right-hand side (profession-oriented) is currently the most important, though this may change over time.
At this point in time, only 5% of cancers in the U.K. are identified by screening. If we focused on improving population health through AI-driven screening, it would take a lot of work to get that up to making the impact of some of these other areas.
By all means, we should be working on population health, but the immediate payoff is on the right-hand side of the framework. If we can get better image recognition, so that overworked radiologists can actually see more chest x-rays and CT scans and be more accurate in doing so, or if we can get a better selection of patients in general practice, that’ll be a quicker hit.
Full recordings of our February 8 and March 10 Revolutionizing Healthcare sessions
Join our community of clinicians!
If you’re a practicing clinician would like to be part of our ongoing discussions exploring the future of machine learning for healthcare, please sign up for upcoming Revolutionizing Healthcare sessions via the link below.
Tim Schubert
Tim Oosterlinck
Andreas Bedorf
