Ask Away! Blog Series Featuring Dr. Michael Hogarth
Welcome to our Ask Away! blog series where you’ll get an opportunity to ask any burning questions to a member of the UCSD ACTRI team and get expert insights.
This month, we are delighted to feature Dr. Michael Hogarth, a distinguished Co-PI of UC San Diego ACTRI. Dr. Hogarth brings a wealth of knowledge and experience as a board-certified Internal Medicine physician and faculty member in biomedical informatics. His expertise extends to his role as Chief Clinical Research Information Officer for UC San Diego Health.
Prepare to delve into the world of research, medicine, and informatics as Dr. Hogarth tackles your most pressing inquiries. Stay tuned for a captivating conversation that will leave you enlightened and inspired.
Let's jump right in and ask away!
Why did you choose to join ACTRI and ultimately become a co-PI? How does this decision fit into the path you’ve taken and want to continue?
Over the years, while faculty at UC Davis, I developed many ties to UCSD. I had active collaborations with several UCSD colleagues (Dr. Barbara Parker, Olivier Harismendy, Lucila Ohno-Machado). I was also part of the UC Davis CTSI and had met Dr. Firestein and knew his excellent reputation as well as that of the ACTRI... Also, a former student of mine and longtime colleague, Dr. Longhurst, had been recruited to be UCSD Chief Information Officer. In 2016, they all called and asked if I would be interested in coming to UCSD to be the institution’s first Clinical Research Information Officer. I jumped at the chance. On becoming a co-PI on the ACTRI grant, for the past 20+ years I have worked with Dr. Esserman (UCSF) to develop a number of innovative informatics systems to support breast cancer research. As an example, for ISPY2, my team worked with Peter Barry at MD Anderson on a novel adaptive randomization engine which we integrated it into the ISPY2 research platform. Additionally, we were the first to build a clinical research infrastructure on a cloud environment – we built ISPY2’s system on the Salesforce cloud. ISPY2 was awarded an innovation award from the NCI. Another example was creating the first clinical research iPad application, which was used in the Athena Breast Health network, a consortium involving the five UC health systems. Through those experiences, I got to see first-hand many challenges clinical researchers face as they conduct research. When the role of co-PI of ACTRI was posted, I thought it would be a great opportunity to not only work with a broader group of talented people at ACTRI, but to also have a hand in our institution’s clinical research strategy over the coming decade or so. In addition to continuing to work with Dr. Firestein and our great ACTRI team, I get to work with Dr. Davey Smith, an exceptional clinician, researcher, and organizational leader. It really is a wonderful opportunity for me to expand my horizons beyond biomedical informatics.
In your own words, what is biomedical informatics and how does it contribute to healthcare and medical research?
That is an excellent question as the discipline of “biomedical informatics” is often not well understood and descriptions can vary significantly even among those in the discipline itself. I will share what I wrote in a recent report in which I described the discipline and its relevance to our institutional missions.“Biomedical informatics is a distinct scientific discipline with a nearly half-century history of scientific endeavor and with significant impact on society. The origins of the field can be directly traced to a seminal article by Ledley and Lusted (1959) in which they outlined how to implement computerized probabilistic clinical decision making. Although the discipline employs theories and methodologies from data science and computer science, it is a unique discipline requiring a blend of scientific knowledge directed at solving problems in healthcare delivery, biomedical science, and medical education. These areas have unique qualities that are often barriers to effective interventions and require multi-disciplinary specialized knowledge for effective solutions.” It is not unusual to view biomedical informatics and health-focused data science as versions of the same discipline. Although they are similar, data science today is focused primarily on algorithms and techniques to analyze data to extract insights, where biomedical informatics adds information driven interventions at the point of care and evaluating these with rigorous experimental methods.
Can you provide examples of how biomedical informatics has been successfully applied in real-world healthcare or research scenarios, perhaps where you yourself were part of the research or deployment of technology?
Great question. There are too many to share, but I can provide a couple of favorites! One of the most interesting for me from both a clinical and informatics perspective was a research project where we wanted to study whether a warfarin dosing predictive algorithm could decrease variations in anticoagulation that typically happens because of its complicated pharmacokinetics. The fact that we did this on a Palm pilot dates me a bit, but many of the lessons learned in WARFDOCS apply to AI-based interventions being attempted today. It was quite an informatics feat to run a mathematically complex algorithm on a handheld device with a low-cost CPU intended for simple tasks like keeping calendars, a simple calculator, and taking notes. We were able to build it and it worked well, so we proceeded with a clinical trial across 6 hospitals spanning 4 different healthcare delivery systems in northern California. A big lesson for me was the impact of “center effects” on outcomes of ‘informatics interventions.’ Center effects are differences in hospital workflows, organizational structures, and clinical/practice culture. It taught me that the ‘real world’ performance of predictive algorithms is heavily influenced by factors outside the algorithm itself. This is why it is so important to test predictive algorithms, such as AI-based systems, in one’s own healthcare institution and not assume it will work as it does in another healthcare system. Another example in which I was involved was the Athena Breast Health Network, which is still ongoing and has a goal of collecting information to drive an automated breast cancer risk prediction as part of routine clinical workflow. The prediction is delivered to a radiologist reading a mammogram as well as the patient’s primary care physician. That infrastructure has also been used for the WISDOM study, which is comparing risk-tailored breast cancer screening compared with current practice. To drive the data collection in Athena during the patient’s mammography visit, we needed to make sure the Athena questionnaire system was aware of a person’s mammogram schedule. If the scheduled changed, the system would need to adjust itself accordingly. We interfaced with EHR scheduling systems using a health data exchange standard called Health Level - 7 (HL7), which then drove a Salesforce built application to send invitations to complete a mammography intake form (questionnaire). To do this, we had to standardize the form across the institutions – at the start, we were a bit astonished that there were 18 different intake forms in active use across the 5 UC academic medical centers. It took 6 months for all involved to agree to a single standard one. We supported intake form completion online prior to the visit or in the mammography suite waiting room during the appointment. We drove both by representing the ‘questionnaire’ questions and skip and branch logic in a standard format (eXtensible Markup Language - XML) used broadly today to share patient records between EHRs. We developed an iPad application that rendered the intake form for those doing it in the waiting room. Consistent with doing real-world based research, we had to make sure the waiting room form could be navigated successfully by someone without computer experience and to complete it in less than 5 minutes without leaving answers blank! It was an extremely useful experience in understanding how people interact with technology while also dealing with complicated integrations that to many looked simple but were multiple systems interacting in milliseconds to enable a cohesive action across them. I am very thankful to having had these kinds of experiences, which were not only interesting from a ‘real world’ informatics perspective, but were helpful in supporting important clinical research.
What are the current challenges and opportunities in the field of biomedical informatics?
The challenges and opportunities are constantly changing in many disciplines, but for biomedical informatics, these lately seem to be titanic in impact and happening at light speed. For example, a few months ago only a few of us had heard of “ChatGPT” and “generative artificial intelligence” or “large language models” (LLMs). In only a few months, they have become the center of attention in the field, with a large fraction of my colleagues doing experiments to understand various aspects of LLMs and their optimal application across a broad range of uses. In concert with this, many of us are actively contemplating the implications for not only biomedical research, and clinical care, but also the training of the next generation of clinical practitioners as well as researchers. A big evolving question is how should we train the next generation of clinicians and researchers in “the age of AI”? Should we incorporate AI assistive systems into the classroom so trainees understand their strengths and weaknesses? If we incorporate these into the classroom and other training venues, will it lead to “clinical deskilling” of doctors? Is it worth risking deskilling in exchange for the bolstering of their “clinical acumen” through the use of AI such that they achieve “master clinician” level? In some ways, we dealt with this in the past decade as EHRs incorporated “templated” orders and notes. This improved safety gained through standard order sets but had to be weighed against the risk that trainees would no longer need to generate all the orders themselves, which is often part of the cognitive process of decision making. If the trainee no longer needs to generate all orders “de novo”, do they also lose some ability to think systematically through those decisions themselves? Another aspect of AI and generative systems like LLMs is how they will impact clinical research. The LLMs, by design, are ‘creative’, however, that also means they can create nonsense. More alarmingly, that nonsense can often be presented in an authoritative manner making it appear as if it were legitimate content. These are often called LLM “hallucinations”, which is an anthropomorphic label that I think misrepresents the situation. I think it is more “confabulation” than anything else. It is also a side effect of how LLMs produce content in response to input. The next generation of LLMs will have “fact checking” modules that will redact these ‘creative additions’ prior to being given to the end user, making the systems less prone to incorrect output. Another increasingly prevalent theme is the integration of research and clinical care at the point of care. Historically, clinical trial management systems (CMTSs) were built in a time when institutional EHRs were uncommon. Today, EHRs are ubiquitous in US healthcare, particularly among healthcare organizations that also conduct a significant amount of research. The HITECH Act was extremely successful in spurring adoption, and after 25 years of having EHRs but very limited adoption, the incentive program moved us from 28% adoption to 96% adoption in only 10 years. Because of this, the ‘cheese has been moved’, so to speak, when it comes to clinical trial management. From an information management perspective, it seems increasingly beneficial to co-locate clinical research information and clinical care information for the same person in the same system. This is even more compelling when one considers pragmatic trials, which often have the same design and information needs as quality improvement interventions prevalent in learning healthcare systems like UCSD Health. Laura Esserman would often say to me “Mike, we should learn from every patient and we should build our systems to do that”. The corollary is that EHR systems today should ideally be designed to indeed do that, whether the patient is part of a learning healthcare system quality improvement endeavor or a prospective randomized clinical trial. This is not a trivial task as EHRs were not originally designed for this, and CTMS systems were not designed for clinical care. This ‘evolution’ is currently underway, making “living through” that transformation an interesting experience and something I very much enjoy – it is like navigating a new challenging hiking trail, which is always fun as well as a challenge.
What technologies or tools are commonly used in biomedical informatics research and practice?
Clearly, one of the top contenders is the electronic health record (EHR) system. In 10-years (2011 to 2021), because of the HITECH act and its incentive program, US adoption of EHR systems increased dramatically from 28% of hospitals to over 96%. This fundamentally changed the underlying sourcing and management for clinical information. It not only changed clinical practice but also enabled digital collection of research related case report forms at the point of care. An added dimension of having EHR adoption has been the ability to more easily introduce and validate informatics interventions. This is the essence of biomedical informatics. Prior to having high EHR adoption, one had to build these informatics interventions as tools that were essentially stand-alone applications with the user re-typing things like lab results for the algorithm to work. The WARFDOCS system I mentioned previously is a good example. We built that before EHRs were widely adopted and we did it on a Palm Pilot. Warfarin doses given and resulting INR lab test results had to be manually entered by the user with dates and times for the system to render a prediction. Today, we would have built that system such that it would pull this data in real time from the EHR, produce the prediction, and show that to the end user in smartphone application. In some ways, the high adoption of EHRs has created a potential virtual laboratory for biomedical informaticists who want to innovate with new algorithms and test them in a real-world context.
We read an article about a study you were involved in that compared written responses from physicians and those from ChatGPT to real-world health questions and understand there’s been widespread speculation about how advances in artificial intelligence (AI) assistants like ChatGPT could be used in medicine… How do you think machine learning or artificial intelligence is contributing to medicine, and what are some notable applications?
When asked about Artificial intelligence in medicine, I often remind folks that AI algorithms have existed in medicine for over 50 years. There have been advances in AI, computing power, and computerization at the point of care, which have now coalesced to bring increasingly useful AI as assistive systems to the point of care. Notably, the first paper published on the notion of computers and medicine in 1959 by Ledley and Lusted specifically highlighted how clinical decision-making benefits from a Bayesian probabilistic approach and how computers could assist us in performing those calculations at the point of care. Interestingly, although the math behind that approach is quite simple, and we now have computers at the point of care, we still don’t see widespread routine use of a Bayesian decision making assistive system. This is likely because it needs correlative causal data, which we are gathering but is still not optimally curated into a “knowledge base” that can be semi-autonomously maintained. I think AI systems can assist in many areas, including this specific task of harnessing EHR ‘real world data’ into a reliable predictive knowledge base that future clinicians will incorporate into their daily workflow to optimize decisions in a patient-centric way. The other thing that has kept AI ‘on the shelf’ so far has to do with socio-technical aspects of using computers in medicine. We should keep in mind that EHRs have only really been ubiquitous at the point of care in the US for maybe 7-10 years now. That means that the majority of practicing clinicians are new to computers at the point of care. This is a practice cultural phenomenon that takes a few years to change. As more “generation Z” and “millennials” enter medicine and become clinicians and researchers, the use of computers will be a natural extension of their work process. Today, ‘computers in medicine’ is still a novel concept in many practices. In addition, even the EHR systems we have today are quite crude and fundamentally mostly word processors in terms of functionality. They don’t have “smarts” yet. So, it will take some time for newly trained clinicians who have ‘grown up’ with things like GPT and AI assistive technology to become the majority of the practitioners. At that point, nobody will question the use of AI in medicine – it will have become part of the “DNA” of medical practice and research.
Dr. Hogarth, you are always using stories to communicate messages and explain challenging concepts. With a field like BMI, which is very technical and somewhat indecipherable to many, how have you managed to use this technique to better connect with your audiences?
I became interested in computers after a work-study internship job I had at the Stehlin Lab in Houston during college. I was asked to work out how to automate a system using one of the first Intel based desktop computers made available. I had to write software using something called “assembly language”, a very low-level language. I was a bit intimidated at first, but in reading the engineering manual about that computer, I found it was not as mysterious as I had assumed. I discovered that modern computers are fundamentally and quite simply adding machines that do this with a binary number system, which is simpler than the decimal system we use ourselves every day. That gave me a great perspective that is still useful today, namely that once you break down complicated technical things to their basic components, they really are quite simple in design. They only appear to be densely complicated, however, they are always an assembly of very simple components. This is the case for high powered CPUs in your computer today to quantum computers and machine learning neural networks. At the fundamental component level, I have found them all to be quite simple to understand. Their complexity comes from layering and assembling them into multi-component systems, but their foundational elements tend to be quite simple. I typically dissect technology into these simpler parts, and that is the approach I take in explaining these systems to others. I start with the simple parts and describe the complicated systems through them. It has tended to work well for me. I think it is true for many things, including biomedical science, that learning “first principles” is the best way to really understand how things really function, and that it is often not as mysterious as once thought.
How do you balance all of your responsibilities? Tell us about your work/life balance. As we recall, you are quite the mountaineer…
It is difficult as we are pulled in many directions in our professional lives, but when I have time to get away, I try to go to places where email cannot follow. I like to get away on multi-day hikes with my oldest son, who is an avid outdoorsman. I also very much enjoy going to the safari park, LEGOLAND, and the zoo with my wife and younger son. Walking the zoo is great training for hiking SoCal peaks! I wouldn’t call myself a mountaineer, but maybe an increasingly experienced hiker. Prior to coming to UCSD, I was not aware of the amazing hiking and outdoor opportunities in southern California. One always hears about the beach and surfing but very little about southern California’s mountains and peaks. It has been a wonderful process of discovery for me to experience the ‘southern sierras.’ So far, I have done most hikes in San Diego, including the Fortunas, Mt Woodson (the ‘potato chip’), Black Mountain, and El Cap/El Cajon, Dos Dios, Double Peak, etc. I have also succeeded in summiting Baldy, Taquitz, and a few other southern California peaks. I tried Mt Whitney via the mountaineering route last June but the weather did not cooperate so I did not get up to the top. However, it was an amazing experience with ‘bouldering’ and ‘slab walking’ and meeting many an intrepid marmot along the way. I look forward to trying more peaks in the near future.
What would you like the research community to know about you?
That I really enjoy being part of UCSD, and that I am always available for discussions on innovative informatics and technology in clinical research.
Don’t forget to email us at: researchcomm@health.ucsd.edu or v3chavez@health.ucsd.edu if you have any questions for him as we will be talking to him again on Instagram Live in the coming months.
Follow our Instagram account for more information here.
Lastly, this Ask Away! Series will be posted on a monthly basis, so feel free to send any questions you might have about research and science our way.
Until next time,
UC San Diego ACTRI