Focused On Precision Data – an interview with Louis Culot and Qi Wei, Leaders of the Philips Oncology, Informatics Initiatives1,2
Precision Medicine is focused on generating and using precision data to develop and deliver drugs for the unmet needs of patient populations who do not respond to the standard of care. Philips approach to precision medicine is to arm care teams to make decisions efficiently, collaboratively, and accurately with expert clinical guidance and a holistic view of a patient’s genotypic and phenotypic information.
Recognizing the growing need for technological advancement in comprehensive xoncology care, Philips has built its Oncology Informatics and Genomics business on Philips HealthSuite with workflows for pathologists, oncologists, and researchers. The platform streamlines collaboration, offers a comprehensive molecular picture, and delivers actionable reports for a better path towards precision care. Philips vision is to become a leader in innovating precision medicine solutions that connect areas such as pathology, genomics, and molecular phenotyping molecular for personalized therapy decision making.
Tolearn more about this platform, we connected with Louis Culot, General Manager, Oncology Informatics and Genomics, and Dr. Qi Wei, Genomic Subject Matter Expert. They have kindly agreed to discuss the program and its aspirations with us.
Q. How does Philips define precision diagnostics and precision medicine?
A. Louis: If I start with the concept of a precision diagnosis, we think about getting a first-time right diagnosis for cancer patients, including doing better at localization and characterization of the disease. Localization is often done with imaging, but also sometimes through integrating imaging sources.
We have examples where we combine MR imaging with ultrasound images to provide a better, more accurate, localization picture. In terms of characterization of disease, we think about doing better by integrating information across modalities such as molecular genomics and imaging features.
And then in terms of precision medicine – that’s the other side of this – how are we treating the patient? You could think about it from a purely pharmaceutical perspective, but you could also think of it more broadly in terms of other interventions. We can think about precision medicine to include image-guided interventions such as ablative techniques and so forth. That’s how we think about precision medicine and precision diagnostics at Philips.
Q. Could you please tell us about Philips Oncology Informatics platform, including its mission and goals.
A. Louis: We think about our platform across four pillars: the first three are localization of disease, characterization of disease, and providing therapy guidance and planning. We’re not a pharmaceutical company, but once you’ve done the right first-time diagnosis, we do think we can inform how you might take the right next step.
And then the fourth, which just as important, is doing this in a closed loop framework. Our customers can perform analytics on their populations and learn from the flow of patient cases through the platform. If you want more detail around that, you can start with the localization and characterization directed to integrated diagnostic. And then once you’ve done that, which, to be honest, is a hard and substantial problem in cancer today, then you can think about the right next step.
By way of example: we’ve included game- changers such as Dana-Farber oncology Pathways on oncology pathways. We also aim to help better matching of patients to clinical trials and increase accrual of patients on trials, and then learning from the population.
Q. Precision medicine, and in particular, precision diagnostics, depends not only on having large data sets but also data set design. Could you please comment on Philips strategy for designing and building its database?
A. Louis: Data model design is often a missing feature in discussions about precision medicine databases. Many initiatives have put big data sets together without thinking through necessarily how to compare “apples to apples” across different systems, different ways of capturing patient information and so forth. We have spent a lot of time on foundations in the data model itself, how we can think longitudinally, how we think about the biology and the medical side of the equation so that we actually have a resting place for data itself.
“Getting the data model right in the first place is a hard thing”
We’re also involved in standardization initiatives to make sure that we’re compliant and even helping shape some of that discussion. Getting the data model right in the first place is a hard thing.
We face a continually evolving challenge because as we learn more about cancer, as we learn more about the biology of cancer, as we learn more about how to treat cancer, these data models are, by necessity, going to need to continue to evolve. Furthermore, different groups in different initiatives are trying to target it from different angles. As such, the data model itself, I think will continue to be a work in progress.
Q. Experience has shown that even the most diverse collections of data can be sparse in certain areas. How will Philips ensure a dense database? For example, will you generate your own data if there are holes, should there be any (especially in light of your point of an evolving model)?
A. Louis: In terms of the data itself, I try to start with the question we’re asking of the data – this is such a complex space, you have to start with the questions. This concept of “we’ll pull all the data together and then we’ll be able to do such great stuff with it” is a little naive. You may get some low hanging fruit out of that sort of an approach, but to drive insights, you have to start with the questions you’re trying to ask. In so doing, you’ll uncover any holes in data that you may need to answer those questions.
For example, we want to answer the question ‘what is the variation in care throughout a cancer network in a certain disease area, like lung cancer or breast cancer?’ We can think about variability and care; understand how patients are being brought in; how they’re being diagnosed; how are they being navigated; and what treatment decisions are being made. We need granular information about the patient population because some variation would be perfectly expected and warranted … and, potentially, some variation that would not be expected.
We need a framework to think through the question in terms of database design (and any subsequent holes in datasets). Along these lines, we do sponsor work to produce data, and we have a very active research program where that data is going to help us answer many questions. In a lot of these cases, we also receive data from health systems that are running the platform and trying to answer these kinds of questions for themselves. In those cases, we make sure that we are acquiring the right data for them to answer their operational questions.
So, this is step one starting from the question side. Then, in terms of how we fill the holes, sometimes it will come from the health systems themselves; sometimes it will come from us sponsoring research; and sometimes it comes from public datasets. We use all three of those sources.
Q. Are your partners willing to share aggregated data among themselves or do some maintain control of it? And the follow-up – are there any restrictions on Philips use of data?
A. Louis: Our general philosophy is to be a custodian of data. There are cases where we do need data to drive the platform, but our general philosophy is that the customers are in control of their data. Some of our customers do want to share data amongst themselves and we enable that. We don’t claim any rights into their datasets beyond what their willing to share with us, and we don’t monetize their datasets. So it’s really within their control. We find some customers want to share data and some want to keep their datasets private.
On this point, though, I’ll say there’s a symmetry with the customer base. We may have a customer who actually wants to share some of their data to collaborate and to bring big data sets together. But then the customer may have other elements of their data, whether operationally sensitive or data that they want to keep to themselves, and we allow that. Again, we have the philosophy of data custodian, and we adhere to that philosophy as a company.
Q. I noticed in online posts that your Genomics platform has a capability to automate pathologist workflowsand to generate customized reports. As this is core to where the field may go, could you please describe the process and details about producing these customized reports for healthcare providers? (see Figure 1 for detail)
A. Louis: One thing I want to ensure with the data and reports is that we capture the connection between the pathologist, the pathologists workflow, the production of these reports, the oncologists, and – the final link in the chain – the clinical application of these insights into patient care. That’s been a missing link in the field and something that has held back more rapid adoption of precision medicine.
Figure 1: Genomic data enter the healthcare continuum (molecular pathology path). IntelliSpace Precision Medicine Genomics enables the implementation, integration, and scaling of precision medicine program to improve patient outcomes.
When we start in the pathologist’s world, we first think about sequencing or next generation sequencing output. These activities need to be run in a controlled environment because these are typically billed as customers’ tests in the US under the College of American Pathologists (CAP) guidelines. They need to go end-to-end in a controlled way that fulfills the regulatory environments as you go from next generation sequencing output through bioinformatics analysis to analytic detecting of variant. And what I mean by analytics, is the mutation here or not? It’s like a biological reality. Do you have a KRAS G12X or do you not? The next phase of this is the functional annotation of it. Let’s say you have a variant – what’s the functional effect? Is it activating the gene? Is it a loss of function? Is it unknown in terms of what it does?
“When we start in the pathologist’s world, we first think about sequencing or next generation sequencing output”
We draw on many public data sources especially as these sources become more standardized and validated by the laboratory (including annotations from publicly validated embedded sources). In fact, laboratories themselves will cross validate our end of the section, which is informatics; but of course, it is all connected back to the sequencer and the chemistry. They have to do an end-to-end validation to be sure that what they see is real
Next, there is the clinical effect or the clinical implication. For example, given you have a KRAS G12X and it has a certain functional effect, what is the clinical decision an oncologist is going to take? Are there clinical trials available for that patient? Are there therapies available for 35 that patient – either FDA approved therapies or off-label indications and so forth? That’s the link into the oncology world that has been missing.
Although we certainly can generate the report in the lab and have the molecular pathologist sign out the case, there is this connection into the oncology world where a lot is dynamic. For example, a patient is sequenced on a Monday and their oncologist looks at it a few days later – in those few days, a clinical trial may open at the hospital. This happens all the time; we find out there may be clinical trials running in cancer care that may not have been available at the time that the sequencing was run. Being able to do these things dynamically and connecting this to the oncology side is really essential. That’s a big part of what we’re doing.
Q. Are these reports shared with patients?
A. Louis: My understanding is that they are part of the patient medical record, a molecular pathologist signs them out, and they would go into the patient’s electronic medical record file, available to patients by request. Some places have portals but, in any case, I believe patients have access to that data by regulation.
Q. To change tack slightly here, Philips focus appears to be currently on cancer. Will Philips eventually work in other therapeutic areas?
A. Louis: That is an interesting question because if you come at it from the outside world, people would say our focus is probably even more heavily in cardiology. Certainly, cancer is a huge focus for the company given the importance of in vivo diagnostics and what we can do in image-guided therapy and informatics.
However, the company is active in a variety of health spaces. As additional diagnostic information becomes relevant and actionable in other spheres, we are absolutely going to be playing a role. Cardiology, sleep and respiratory in the acute care space – we are constantly looking across health spaces where we can really have clinical impact. Some of these areas are more research focused and we are paying attention to those. We are active in some research programs, too. As they move more into clinical use, we are going to be there.
Q. Do you see the methods, tools, and analytical capabilities you developed for cancer as extensible to other therapeutic areas?
A. Louis: Everything on our platform across the HealthSuite is built to be extensible. I even don’t like the phrase “disease diagnostic” because it implies you are trying to be everything to everybody. However, we do build it to be extensible across different diseases. Just as we built out our capability in oncology, we actually have built a capability for other disease areas using data we are collecting from imaging equipment or other tools currently used to guide therapy. As molecular features become important, we will incorporate those on the back of the same data models and the same principles that we are using here.
Q. Do you want to make any comments about your analytical platforms? For example, are you using machine learning or artificial intelligence?
A. Louis: We use any and all analytic techniques. It does come back to the questions we are asking as we try to generate hypotheses, and then decide on what we can do with that information. We have a program in molecular pathways diagnostics, which is based on a biological insight moving from gene expression that was built with a Bayesian model framework, as an example. And because that would be potentially used as a diagnostic at some point, those models have to be frozen so they can’t be continuously changing over time.
In short, we use all of these techniques on the datasets to generate new models, predictive models, or insights. And then the question becomes, are they dynamic, can they change day-to-day? It really depends on the application. If the technique is used operationally, the technique must certainly be dynamic. If it is used more from a diagnostic perspective, then there is more regulation around it in terms of how it has to be characterized so that the regulatory people can work with us on validation, and then things cannot be as fluid in those models.
“Cancer is an extremely complex disease and extremely complex to treat. So, we’re not trying to shortcut that process, we’re trying to enable it.”
But in terms of techniques, yes, we certainly use machine learning, deep learning, Bayesian networks as well as descriptive and predictive analytics across these datasets. I would say that we need to come back to the data and the data model – these are the more essential ingredients to all of this versus the methods themselves. Because these methods are (or will become) more or less widely available.
Q. I think your point being that someof these methods are more applicable in certain areas depending on the question you’re asking. So that was helpful. I appreciate the distinction you make between precision and personalized medicine. Can you expand on that a bit?
A. Louis: Yes – so again, we start with this concept of precision diagnosis and bring together different sources of information – all that information currently maps to a set. That gives us a better insight into the patient as a step one. We’re not yet at the point where we design a medicine based on an individual person. There are maybe some exceptions to this – for example, engineered immune cell techniques for an individual – but those are really edge cases of general practice. Precision diagnosis drives a much more granular approach to how we treat patients.
If we can structure patient data in a more granular way up front based on, for example, molecular features, line of therapy, certain comorbidities, then we’re at the forefront of precision diagnosis, precision medicine.
Within that context, though, you might come up with a set of options for patients that could be personalized. We have examples with some of our clinical collaborators discussing patients who might be concerned about hair loss with a certain treatment; but another treatment option might have a slightly different efficacy profile that is less likely to result in hair loss. Health care providers might, within a set of precision options, personalize the therapeutic choice one way or another.
Doctors make such decisions with dosing as well. If a patient is experiencing side effects, doctors often adjust dose levels. That is personalization. Another side effect in some therapies could be neuropathy that patients may wish to avoid for quality of life. That decision made by the clinician and the patient is, to me, the distinctive nature of the personalized aspect of precision medicine. At Philips, we are not trying to tell doctors what to do, we’re trying to help them think about best practices for their patient and what options might exist. Then they can work with their patients on personalizing that. That is how I think broadly about the complementary aspects of personalized and precision medicine today.
Q. This Journal has adopted a working definition for precision medicine of creating precision tools and drugs that a physician would then personalize with the patient in the office. In effect, we agree that the personalization side is done by the doctor and patient.
A. Louis: Right, that’s right. And one of the things we’re trying to do, is again, develop this concept around pathways, best practices, and clinical trial matching. I mean this is all in the world in my mind of what I would call precision medicine. But then moving to that next phase, some may think that narrows the choices, but we really haven’t, we’ve really expanded the choices, done under the concept of what would be best practice. Then it’s really between the doctor and the patient and their knowledge to personalize the treatment for them. And the choice could be one of the appropriate best practice treatments within that set. It could be something different for some rational reason. It could be dosing distinction. There’s a whole lot of variation here. As you know, it’s extremely complex. Cancer is an extremely complex disease and extremely complex to treat. So, we’re not trying to shortcut that process, we’re trying to enable it.
Q. Thank you very much, Louis. I’m going to switch to Qi now. Qi, could you please tell us about your role regarding reviewing molecular diagnostics and other responsibilities?
A. Qi: To give you a little bit background … I came from the trenches in the clinical arena as a board-certified Molecular Pathology Lab director with close to 30 years’ experience in molecular pathology. I spent my last three years with the Miami Cancer Institute at Baptist Health South Florida, where I established a cancer genomic profiling program. I served as a director of genomic lab services and it really gave me a unique perspective on how cancer genomic profiling results are used in the clinical setting and how precision medicine is being practiced.
Because this is a new position, both for Philips and myself, my role as a subject matter expert is still evolving. Primarily, I see myself as a bridge between our Philips solutions and the customer. More specifically, my role is to help the clinical customers, and at the same time, take feedback from customers to improve our platform.
Q. Along those lines, do you work directly with the collaborators? For example, have you found yourself providing advice on experiment design or assay types? (see Figure 2 for detail)
A. Qi: Yes, as a part of our solutions, which provides consultative services for the genomic and molecular pathology labs, we provide the labs help if they need it for their assay and validation plans. It’s part of the end-to-end service that Philips offers.
A little bit of the details of what we do with our Genomics solution: first, we facilitate the clinical NGS assay validation based on well- accepted NGS validation guidelines from CAP, the College of American Pathologists, and, second, we discuss and establish the validation plans with the laboratories. Finally, in our role as bioinformatics solution providers, we generate the analytical validation documentation.
Again, we work with the customer in analyzing all the data. So, we provide the lab with the assay specificities, sensitivities, the limit of detection documentations. At the end of the lab validation, we provide the summary of the validation results and the details for each stage of the validation, the whole nine yards, to the customer. At the end of the day, we are providing informatics solutions through our platform.
Figure 2: Philips’ Intellispace Genomics an open annotation framework offers a genomic knowledgebase that leverages literature reviews and clinical information.
Q. And how does your work with collaborators inform Philips platform? That is, does Philips use data from clinical customers to revise and refine the platform?
A. Qi: As a solution provider, we work with many different types of customers at genomic labs – some labs are focused on the hematologic malignancies, some labs focus on the solid tumors. So, it depends on what customers are offering. In the solid tumor arena, most of the genomic data or genomic profiling results are used for the therapy matching and the clinical trials.
For hematologic malignancies, in addition to the therapy matching, most of the results are analyzed to aid the pathologist for the diagnosis and the prognosis of the disease. By working with the labs, we see the differences, the small nuances from each arena. So, we bring this feedback from the customers and incorporate it into our platform designs.
Q. Could you also provide some background on how Philips patient data could enable clinical decision support teams? How do they use the reports for their work?
A. Qi: We pick up after the lab has done their sequencing. The sequencing analysis steps usually consist of two steps. The first step is called the secondary analysis and follow up with the tertiary analysis. In the secondary analysis, the tumor sequence is aligned with the human reference for the subsequent variant annotation.
The tertiary analysis follows the secondary analysis; the tertiary analysis is what we refer to as clinical annotation or “interpretation”. The core of tertiary analysis involves the biological classification of the detected variants, annotating the clinical relevance of these variants; determine the clinical action-ability in term of treatment options and availability of clinical trials. As the adoption of NGS grows in the clinical and molecular diagnostic arena, the need for accurate and timely interpretation will become even greater. Most of the genomic laboratories are not well equipped for this. Philips platform serves all stages of NGS data analysis, including tertiary analysis.
We are utilizing several clinical knowledge databases including Jackson Laboratory Clinical Knowledgebase (JAX CKB) and MD Anderson Precision Oncology Decision Support Knowledgebase (MDA PODS).
Philips platform supports customizable and standardized reporting follow the CAP guidelines, while also makes reports easy to understand and act upon by the medical oncologists and other providers. The reports usually contain the FDA approved drugs associated with the variants within the disease indication and outside the indication and clinical trials as well.
Q. Does Philips have any plans to license rights to its database for clinical IVD or LDT development, either with collaborators or other organizations?
A. Qi: As Louis said before, all the data, the clinical data that the laboratory generated, all that belongs to the patients. We do not monetize the data; however, if the customer wants to share the data, we provide them the ability to do that. We are unlike some big labs and we do not own the data, the customer owns the data.
Q. Do you have any other comments or words you’d like to add, Qi.
A. Qi: I think that with genomic profiling, we’re still in the infancy of understanding and using it in the clinics. We only see the tip of the iceberg. As more clinical trials and the more drugs come into the market, precision medicine is really going to take off. Even these last three years, I see this from the trenches, I see the difference, and it’s a really exciting time.
Thanks to both of you, Louis and Qi. We appreciate your insights.
Qi Wei serves as a Subject Matter Expert (SME) for genomics at Philips Healthcare. Prior to join Philips in October 2019, he was the director for genomic laboratory services at Miami Cancer Institute/Baptist Health South Florida where he successfully established and implemented cancer genomic profiling program. Qi has 30 years’ experience in molecular diagnostics especially cancer genomic profiling for hematologic malignancies and solid tumors. He was the scientific director for molecular diagnostics laboratory at Children’s Hospital Colorado/University of Colorado. He received his PhD in human genetics and undergraduate study in biochemistry.
Louis Culot is General Manager of the Oncology Informatics business at Philips, where his team is leading initiatives to use informatics to better localize, characterize, and guide treatment decisions in cancer, in support of Philips mission to improve the lives of 3 billion people by 2025. Louis’ team uses advanced data modeling and analytic techniques to help clinicians better understand the complexity of the disease made possible through the digital transformation of healthcare, and drive continuous improvement from early detection through treatment and follow-up care.