Introduction - Joanna Grossman, MD, MSc
🫛 This edition of AI, Software, and Wetware features an audio interview with Joanna Grossman, MD, a 🇺🇸 USA-based physician-scientist-innovator specializing in biotech clinical development & adoption. We discuss:
Why only about 50% of evidence-based practices in medicine ever get adopted into routine clinical practice, and why it takes an average of 17 years for them to be adopted after they are published
Overcoming the key barriers that prevent innovative healthcare information technologies, devices, drugs, and approaches to care from reaching patients.
Why it’s challenging to abstract accurate and complete patient-level data from the EHR to evaluate healthcare quality and safety problems.
Developing a predictive analytics algorithm using EHR data with Georgia Tech Research Institute to predict bloodstream infections in pediatric oncology, stem cell transplant/cell therapy patients and identify real-time, individualized, actionable factors to reduce infection risk
How wrestling with US health insurance companies is one of the hardest aspects of delivering high-quality care to patients
Using AI tools carefully (“no trust; it’s all verify”) to help her keep up with the latest in published medical research and increase her productivity.
How she uses AI tools to focus and condense her writing and to structure her work with templates
Living by the “billboard rule” for personal data for herself and her kids
Dealing with phishing attacks and scammers impersonating life sciences recruiters
Her thoughts as a caring physician on how and when patients should & shouldn’t use “Dr. Google” and “Dr. ChatGPT”, why there’s no silly question, and on feeling empowered to get the answers you want and need from your physician
and more. Check it out, and let us know what you think!
This post is part of our AI6P interview series on “AI, Software, and Wetware”. Our guests share their experiences with using AI, and how they feel about AI using their data and works.
This interview is available as an audio recording (embedded here in the post, and later in our AI6P external podcasts). This post includes the full, human-edited transcript. (If it doesn’t fit in your email client, click HERE to read the whole post online.)
Note: In this article series, “AI” means artificial intelligence and spans classical statistical methods, data analytics, machine learning, generative AI, and other non-generative AI. See this Glossary and “AI Fundamentals #01: What is Artificial Intelligence?” for reference.
Interview - Joanna Grossman, MD, MSc
Karen: I am delighted to welcome Dr. Joanna Grossman from Georgia, USA as my guest today on “AI, Software, & Wetware”. Joanna, thank you so much for joining me on this interview! Please tell us about yourself, who you are, and what you do.
Joanna: I’m a pediatric hematologist/oncologist, physician innovator, and strategic clinical development advisor to biotechs, pharmaceutical companies, AI-driven digital health startups, and healthcare systems.
I’m also a mom to two boys who are seven and nine; an equestrian; a problem solver; and a podcast and audiobook junkie. Particularly ones about health, tech, and personal growth.
Karen: Wow, you keep busy.
Joanna: Yeah, it’s a busy life, but it’s fun.
Karen: Okay. Tell me about how you got into this area of work?
Joanna: Sure. You know, I never thought that I would be the person who would develop the next breakthrough drug in a lab. But I always thought that I could be the person who could help people survive their therapy by coming up with innovative approaches to care delivery or technologies that can make care delivery better quality, safer for patients, more efficient, more accessible, and in that way improve outcomes for patients, particularly with cancer, because that’s my focus.
But obviously most of the things that I do translate across specialties to all aspects in all settings where medicine is practiced. And I fell into this specific way of doing this, kind of in a roundabout way. I started out as a real regular translational and clinical researcher. I had gotten my Master’s of Science in clinical research from Emory during my fellowship.
My projects were very much focused at the time on risk stratification and how do we tailor therapies, that we’re not giving too much therapy to the ones who don’t need as much, and giving more therapy to increase the chance of cure for those who need it. And in that way, helping to improve outcomes with fewer late effects for patients, given that I take care of kids and they’ve got a whole life ahead of them. If we are able to do our jobs well and we have a little bit, or a lot of luck on our side usually.
When I finished fellowship, I started doing more quality improvement, patient safety work for my division that I was working in at Children’s Healthcare of Atlanta and Emory. And how I got pulled into it was, one of my mentors, and now longtime colleague, she wanted me to help leverage my data analysis skills and my clinical translational research skills to try and enhance the level of rigor with which we were doing the quality improvement patient safety work. There was not as much data-driven decision-making as we would’ve liked at the time. This is 10, 15 years ago. It still could use plenty of work. But it became really clear to me that a major barrier to doing this kind of work on the local level is actually the ability to get information out of the electronic health record, in order to be able to use that as a basis for designing interventions that can improve the way that care is delivered locally.
Because much of that does depend on the systems that are in place at a given institution and how the operational environment is, and which technologies they have or don’t have, and differing approaches to workflows. So a lot of that. In order to really drive improvement on the local level, you obviously need to be able to have local data. And so, while you can certainly, and I have published and do publish in patient safety and quality, as more of a, “Here’s how we did it, and here’s what we found, and here are things to consider at your own institution.”
But it’s certainly different than research, and it also is different or even less infrastructure than research. And clinical research suffers from some of the same problems, but in particular, these sorts of initiatives on a local level really suffer from lack of ability to get data and use it in order to drive evidence-based change.
So that was my inspiration for going in this direction. One thing led to another, and I ended up collaborating with Georgia Tech on several projects, from devices to digital health solutions, including predictive analytics for bloodstream infections and patients, for example.
And now I’m focusing on advising others on the barrier. Mostly where my real expertise is, is on overcoming the barriers to implementation and adoption of evidence-based solutions or technologies, or other innovations that have data to support their implementation.
Karen: Yeah. I’ve got a couple of questions. You mentioned the electronic health records or EHR. For an audience member, that would be something like the Epic MyChart systems, or systems like that?
Joanna: Epic MyChart is the patient-view side of it. We obviously have the healthcare provider side of it. And actually even within that, nurses have a different view than physicians do. Pharmacists have a different view. So depending on what your role is, there’s different views that you have. And there’s the patient view, which comes with MyChart. So on the providers, on the physician side, on the healthcare team side, that’s where all the input happens. Patients may send messages through a MyChart portal, and those do get added to their medical record. But the majority, all the clinical documentation, orders, results, letters or communications from other specialists that they see, or other care that they’ve gotten elsewhere, that would all be found in the healthcare provider view.
Karen: Yeah, that makes sense. I’m just trying to think about, if someone’s hearing this, how do they relate that to something that they would experience? Or what would their touch-point be on a day-to-day basis?
Joanna: This is interesting because the 21st Century Cures Act required the release of results to patients, in a time so that they could have access to their own medical information. Of course we want patients to have access to their information, but certainly we don’t want them to see a concerning diagnosis or something that will cause significant distress if we haven’t had a chance to review it with them first. So those are some of the drawbacks of that.
For the average person who is looking at their MyChart record, it might have a note from their visit if they’re an outpatient, then their note from their clinic visit that day. If they’re an inpatient, it’ll have the daily progress notes and any notes from consultants. It’ll also have radiology reports and pathology reports and other lab results that they can see. And different practices have different ways of implementing this. So there’s certain information that is accessible via a portal and certain information that still requires patients to request. From as best I can tell, that is very much a system- or practice-dependent decision, in the way that they’ve interpreted the regulations and a lot of times what the attorneys say that they have to do.
Karen: Right. Thank you for explaining that. You mentioned barriers to adoption. Do you find that the biggest challenges are in actually getting enough good quality data to do the kinds of analysis you want? Or is it really more, okay, we’ve come up with a recommendation and we have data behind it, but then convincing the practicing physicians and such to actually believe in it, versus what their experience and their gut tells them?
Joanna: This is one of my most favorite things to talk about! There’s a couple aspects to this when it comes to local data for local improvement activities. For example, if we notice that there’s some sort of a medication or an order is not being released in a timely fashion, or some medicine is being delayed with the administration on the inpatient side, and we’re noticing that that’s happening repeatedly, understanding exactly, teasing apart where that problem comes from, and actually pulling the data to see how often it’s happened.
All hospitals have these online reporting systems. And so when somebody notices an error, it’s not a punitive system. It’s meant to contribute to a culture of continuous improvement and support good patient care and high-quality care that’s safe and effective. There’ll be certain things that’ll trigger that, that they weren’t done the way that they were supposed to be done. But many times, when there’s a problem with that, just by virtue of the fact that there is a problem, there’s a trigger missing, like people are not noticing for some reason, then a handful of people will notice the report it, and then there has to be a root cause analysis and people will get together and try and figure out, try to understand the problem better and pull data to try and get a better sense of the scope and approach to fixing it.
From that standpoint, getting the data to do that is a heavy lift. It’s very difficult to do because the electronic health record as it’s currently designed, was really designed to support billing, not really actual care delivery and communication. And that actually continues to worsen over time rather than improve.
Every time there’s new legislation or change in billing codes, all of a sudden the requirements for documentation end up changing. And they’ve really never become less onerous in my experience. But when it comes to the outcomes of clinical research, the data, it’s not a problem of not believing in the results or not having evidence to support the use of some approach to care, whether it’s a drug or some sort of a care pathway or protocol to manage a given frequently occurring condition or situation seen in a specific population of patients. So that data’s there. It’s just that unfortunately, the major barrier is that many times there’s no standard approach to actually implementing these things, right?
Like if you’re like taking a new diagnostic test, for instance, you can have the best data in the world, and you can have clinicians that believe in it, but if they’re unclear about exactly when to send it in somebody’s therapy, how that process is going to work; if it’s not on formulary at the hospital, then they have to advocate for getting it added. There’s a million layers of bureaucracy when it comes to that. And then there’s all kinds of processes and policies and whatever, depending on what reimbursement is. Really, there’s just countless reasons why innovations don’t get adopted. But evidence or education is not one of them.
And that is something that there’s a ton of published literature about. But still many biotechs and pharmaceutical companies don’t understand this aspect of it. And so when they are planning their clinical trials, there’s not enough thought given to how this is going to work in the real world outside of the study.
So when it’s over and you don’t have the same resources in place, it’s not the same controlled situation then, actually finding a way to implement it without the sort of guidebook of a clinical trial protocol. You’re back to square one a lot of the time unless your institution was really very involved in the development in some way or another. It’s not too dissimilar from things that already exist and there’s already systems in place to address those.
But unfortunately, according to the literature, only about 50% of evidence-based practices ever get adopted. And on average it takes about 17 years for them to be adopted after they are published.
Karen: Wow.
Joanna: Yeah. The field of implementation science arose to try and bridge the gap between knowledge to practice. Distilling the things that we learned from research and innovation and data and having that actually make it to the clinic to benefit patients.
Karen: So are you involved with interacting with patients during clinical trials and studies? And then seeing, okay, this is what worked in the trial, and now how do we make it work in the wider world?
Joanna: Ideally those things are done early on. So in the very beginning, when you’re first doing your early-stage feasibility trials, your phase one trials or your translational work even when you’re just starting to figure out how to bring this into the clinical world from the laboratory or engineer’s office world. That is something that the FDA has started to focus on, is trying to encourage planning for what happens afterward in the context of doing these early feasibility trials. Otherwise, there’s a lot of money that gets wasted on developing things that, while they appear to be exciting and innovative and effective, they just don’t get adopted. That’s usually the goal. You want to be able to plan for these things at the outset.
So actually thinking through, what does care delivery look like in the real world? And how can we design a clinical trial that still has all of the controls and internal validity that’s needed to demonstrate that therapy, the device, the technology, whatever it is, is effective for whatever outcome it’s meant to achieve. If you don’t consider during that time exactly how this is going to look in the real world, not only won’t it get adopted in the way that it should, but the effectiveness, the external validity, like how well do these data translate to what happens when you use the innovation in real life?
When there is a wide gap between how it was implemented in the context of a clinical trial and what happens in the real world, then you’re not going to see the benefit that you expected to have. Because there are complications that happen or patients have to come back to the hospital because there’s issues with the way that something needs to be administered or monitored or. And that wasn’t accounted for initially during the trial because so many things can happen because there’s personnel in place and there’s protocols in place. And so there’s all these supports that are there in the context of clinical trial work that are no longer in place when it comes to adopting it in the real world.
Karen: It’s interesting for me to hear this side of the whole world of clinical trials. It’s been over 20 years now, but I was involved as a patient supporter when my late husband had lung cancer. And he went through multiple clinical trials with different radiation protocols and different chemo drugs that were on trial. Everything was always in a clinical trial scenario for him because there wasn’t any effective established treatment for it. Lung cancer has, I think, still been very problematic. That probably doesn’t affect your pediatric patients so much, but it’s definitely an issue in the bigger world for adults.
Joanna: Well, to your point, oncology, whether it’s adult oncology or pediatric oncology, clinical research is very much ingrained in that practice, and particularly in pediatric oncology. The Children’s Oncology Group [COG] is the largest international cooperative group dedicated to pediatric cancer research. Most institutions who treat kids with cancer will at least treat patients according to the most recent standard arm of the last trial or whatever has been adopted as the standard of care. Whatever the next trial ends up being based on, there’ll be an experimental arm that compares to this to whatever the last therapy is.
The number of trials that, at least, COG has open is quite a lot. And then that doesn’t even necessarily account for the drug company trials, or investigator-initiated sponsored research. There’s all kinds of different ways that clinical trials come to be in a healthcare setting, and different ways that patients can learn about them and be enrolled in them if they are interested, you know?
And I think that when you’re in a field like this, there’s a lot of urgency and a lot of desire, especially for diseases in particular, where there’s either still less than ideal outcomes or significant toxicity from therapy. My initial area of interest was in optimizing therapy so that people are getting enough to cure them, without getting so much that they’re sick in other ways for the rest of their lives. I obviously don’t know how your experience was. Maybe you can share a little bit more about them if you’re willing to.
But, regardless of the phase, there’s a protocol that has to be followed. There’s reporting guidelines, there’s monitoring guidelines, there’s all kinds of things that usually happen in the background. It’s interesting to hear your perspective on it because I know what all goes on. I don’t know if it feels like there’s just all kinds of things. Like if it just feels like, sort of, you’re a guinea pig and you’re not really clear on why we’re doing this, or that, or the other thing. I’m sure there’s some feelings like that. I think that it would be normal to feel some uncertainty when you’re participating in a clinical trial. But also I know that gives people a lot of hope too. I don’t know what your experience is, how it worked in real life.
Karen: Yeah, like I said, it’s been over 20 years, but there’s still some parts of it I remember vividly and will probably never forget. The biggest problem that I remember, with dealing with his clinical trials, was fighting with his insurance for some of the basic things. Like at one point, this was very near to the end of his treatment, where it was obvious that he was dying and that he was in trouble. And we were fighting with the insurance company. He was losing, what, 10 pounds a month, no matter what we did. The doctor said, “Okay, we’ve got to do something to stop the weight loss.” And so the doctor wanted to try anabolic steroids. And the insurance company said, “No, we’re not paying for that.” You know, he’s not trying to lift weights and get into a competition here; he’s trying to live. Fights like that with the insurance were the biggest problem that we had by far. The clinical trial protocols were generally well explained and understood. The doctors were really good. We didn’t really have problems with that so much. I remember the insurance fights more than anything else.
Joanna: Oh, that’s the bane of all of our existence. Even when it comes to standard of care, we still have to advocate for patients and argue with insurance about paying for things that are widely accepted as standard of care. And I personally wasted hours upon days on doing paperwork and doing peer-to-peers and working with case management to figure out what alternative solutions there are that we can do until we can get something for home use for example. They’ll pay for it when they’re inpatient but not outpatient. Or any number of things that really boggle the mind. Especially when you consider the expense of being inpatient versus being outpatient and the cost, whatever it is, right? Some finance person somewhere makes sense of it.
Karen: Yeah, we could probably do a whole episode and go off on that tangent!
Joanna: I’m sure. I’m sure. Yeah, I’m sure. I’m sure we could.
Karen: Yeah. But I do want to get back to the AI and analytics part of it. It sounds like you’ve done some neat work on maybe using machine learning to help develop these data-driven tools and to provide the evidence. So I’d like to hear a little bit more about that, about an example of maybe a tool that you’ve developed.
Joanna: Yeah. So one of the projects that I worked on in collaboration with the Georgia Tech Research Institute was to develop a predictive analytics algorithm that would be able to calculate a risk score for patients admitted specifically for pediatric oncology and stem cell transplant or cell therapy patients who have central venous catheters in place.
For people who don’t know, when you’re receiving therapy, there’s a couple different types of these, which are implanted devices that go into the large blood vessels that feed directly into the heart. It allows us to give medications that aren’t really safe to give through peripheral veins, because they’re too caustic or damaging to the veins, or they need to be given at a rate that’s not safe to give through a peripheral IV. Those devices, as all things that are foreign that are implanted, whether it’s a material for an orthopedic procedure like a hip replacement or some screws that are holding somebody’s broken bone together, or really anything that you weren’t born with, has the opportunity to become infected. And particularly for central lines, in order to actually give medication through them, you have to access them. And they’re very specific protocols about how the lines are cleaned and maintained on a regular basis as well as when they’re accessed and when different components need to be changed out.
And most aspects of the central line protocols are evidence-based. There’s many components, and those things are carried out. They’re executed primarily by the nurses who are caring for the patients. And people are stretched really thin. We’ve had a big problem with several components. There’s a workforce problem if there’s still only one nurse for four sick patients. If everybody’s due for something at the exact same time, well, that’s not actually possible. So when it comes to these routine care delivery things, there’s a practice and educational component to it. So the more senior nurses that we lose, because they’re leaving healthcare or doing other things, retiring on a beach somewhere, the more new grads that we have. There’s nothing wrong with new grads. We need new nurses, we need new doctors, new nurses, new nurse practitioners and PAs [physician assistants] and all the members of the healthcare team are important. People need to come up through the ranks and get experience somehow or another. But when there’s not enough senior folks around to mentor and guide and help, on top of the fact that they’re just, straight, not enough nurses, or doctors for that matter, to manage all the patients safely, then certain parts of these routines that are known to prevent these infections might get missed, might get delayed, might not be done in quite the manner in which they should be done just from an educational perspective. But it’s really more than education because most people know what needs to be done, it’s a matter of the ability to get it done.
The central line maintenance bundle is the term or the name of the protocol that’s used to help maintain these lines on a regular basis, on a daily basis. Things have to be done daily. Things have to be done after infusion of certain products and other products. Different timeframes where things have to be changed out or not changed out, or other things that need to be done with the dressing. Or all these things are meant to prevent infection. And they’re all documented in the chart in some way or another.
The problem is, and this is one of the things that I already understood, but became even more acutely aware of, and disheartened or dejected by, I guess, was realizing there’s really no one source of truth in a medical record. And so when you’re trying to design predictive analytics or really any AI, machine learning tool, you have to figure out: Where is the data coming from? Where are your inputs coming from? Whether you’re training a model or whether the model is performing on your new population, or your patients when it’s been implemented, it has to know where to get this information from. There’s always a problem of: Was it documented at all? Things may be done but not documented. Things may not be done. Things may be done, but not quite correctly or not at the right time. And then, things obviously can be done on time and correctly and according to the protocol.
And so what we set out to do was take all the aspects of this central line maintenance bundle that are performed, because it’s all ordered in the patient’s chart. And when the nurses perform these tasks, they document. There’s a place where it’s supposed to go. But it turns out that there’s some variation in where it ends up – even if it’s being charted correctly, or if there’s a change over time in where they’re supposed to document certain things, or whether there’s an order attached or not attached to it. So there’s all these technology and data input factors when the data about what is being done to a patient is being put into the medical record. There’s a huge amount of variability in where that can be found, if it can be found.
At the same time, there was a lot of work being done on sepsis in adult populations and pediatric populations. There was a whole national quality initiative related to preventing sepsis, which is a life-threatening bacterial bloodstream infection. Requires prompt intervention to diagnose what is the organism that is causing the infection, and make sure that patients are getting treated with appropriate antibiotics for an appropriate length of time.
So when we actually started working on this, these simultaneous initiatives were going on in the attempts to develop the analytics. And I know Epic was actually doing a parallel project at pretty much the same time. And the thing is, like with all interventions, or if you’re going to predict something, the population is critical. You’re not going to predict a central line associated bloodstream infection in a patient that doesn’t have a central line. That’s just the simplest example of why population matters.
Well, in pediatric oncology, there are things in their lab results, for instance, that are associated, but they’re not going to be strong predictors. For example, all these patients who are receiving intensive chemotherapy are going to have low blood counts. A person off the street who you’re trying to use the surviving sepsis pathway and the analytics that they’ve developed in order to try and help more promptly diagnose and treat that in a normal patient, normal being otherwise healthy, maybe a couple comorbidities or whatever. Those labs and the risk factors for their infection are not the same as for a patient who is a pediatric oncology patient with a central line who’s had toxic therapies that not only have effects on organ systems but suppress their blood counts and their ability to fight infection.
But many of the labs are expected in an oncology patient and it’s not a meaningful predictor. Because if everyone’s got a low white count when they have an infection, well, sure. This is a no-brainer. We know that because we caused that with the medicines that we gave, you know?
Karen: Mm-hmm.
Joanna: But with somebody off the street, though, who comes in and they have a very low white count, for instance, but they were theoretically otherwise healthy or whatever, that is a meaningful predictor, because we do know that that’s a risk for infection.
Right. But if we’re trying to figure out amongst all of our patients who are admitted to the hospital, let’s say about 80% of them or more have a low white count, right? And we’re trying. But they don’t all get central line associated bloodstream infections.
So trying to identify who is at risk for those things, given their unique characteristics based on their disease, based on therapy they’ve gotten, whatever other complications they’ve had and how their care is being delivered, for example, from the aspect of performing the central line maintenance bundle.
So what we set out to do was try and first figure out what are those determinants that would make one of one of our patients more likely to have a central line associated bloodstream infection versus not. Let’s say that the idea was to give a patient a risk score of some sort that would push to a clinician. But having a score without anything actionable is like, “Okay, thanks for telling me”. It’s like all the other alerts that we get that we just hit silence on, you know? Because it’s like, “All right, well I already know that the heart rate is high because the baby’s screaming.” I’m just going to keep silencing the machine.
But then you pull these things into an algorithm and are able then to actually have it tell you what were the factors for this patient that contributed to their risk score being high. And then based on that, the idea was for it to select out the things that were actionable. It can tell you about other factors, right? But if it’s not actionable, then it’s just a nice-to-know thing. Let’s say a patient has a risk score that’s a certain way because a dressing wasn’t changed in the timeframe that it should have been changed, for whatever reason. Maybe the patient was in radiology getting imaging done, and by the time they got back it was shift change. And then, before you notice, it’s tomorrow. You know, these are not intentional things where things are not getting done. These are care delivery obstacles. And the point was to try and identify when those things are happening so that it can be addressed before it becomes a problem.
We had developed the algorithm on a training sub data, like 3 million data points over several thousand patients. And we got to the point where we were going to need to figure out how to integrate so that we could do some prospective validation. And that is where it went to die, that project, unfortunately. I’m assuming many of your listeners are familiar with the application integration problem within healthcare, but actually being able to have software that was not developed internally by Oracle or Epic or whoever, it’s very challenging to get willingness to make that happen, to remove the barriers to make that happen. People will always say “cybersecurity concerns” or “health IT concerns”, but when you’re deploying something within the infrastructure of the organization, that shouldn’t really be the fundamental issue.
But nonetheless, it just sort of went to die, which was unfortunate because it was well over a hundred thousand research dollars. We had grant funding that was supporting that project, and had the opportunity to really impact care delivery for patients. This was back in 2018. I’m sure now, with some of the programs that have been developed by the individual EHR company providers, maybe they would be able to perform some of these things. But there’s not the same subject matter expertise that goes into designing these things.
And so it’s the same reason why the surviving sepsis initiative doesn’t apply to all patient populations. These things have to be applied into the population in a way that is appropriate and takes into account their specific risk factors and disease characteristics and things like that, which is not something that computer scientists would know. I mean, because why would they, right? That’s not their area of expertise, right, any more than it is my ability to code a machine learning model. So it’s something that I comment a lot about on LinkedIn, about the importance of having a clinical collaborator for when you’re developing any sort of digital technology. For so many reasons, not the least of which is making sure that you’re considering the domain knowledge, right, that is required to ensure that you’re even targeting the right things.
After that, it then also becomes an adoption implementation problem. We’re back to that again. Where even if, let’s say you did do that or you’ve managed to figure it out, and you managed to accurately figure out what those determinants should be, what those data should be, when it comes time to deploy, if you haven’t considered how the workflow is, or exactly understanding where even. Our technology was designed based on where data’s input in our version of Epic, for instance. That doesn’t mean that every institution is going to input that data in the same place. Many of these companies tout as a benefit this ability to customize and whatever. There’s ways to do customization. People talk about headless EMR, EHRs, where data just basically lives in a pool, and then you get shown different views of the data based on what your role is. It’s the user interface and the information that you get. It’s not put in a different place. It’s all in the one source of truth. It’s just delivered to you in the way that you need to see it, you know? That’s sort of the dream, because that then enables interoperability and enables all kinds of analytics and data capture and opportunities for innovation that are currently impossible because of some of the infrastructure of the software.
Karen: Yeah, again, we could spend a whole lot more time on so many things! Like you mentioned the difficulty of getting the data and then getting it charted. There’s all these innovations with medical scribes and trying to offload people from recording the data and, but they can be error prone. And then there are the genetic services where there’s genetic data and some signs that’s being pulled in. There’s so many other directions we could go.
But one thing that you had mentioned in our earlier discussions was about using it not just for your research work and for developing the tools, but also for looking for relevant papers and articles about the subject. Can you say a few words about what you do with that, how you use that?
Joanna: Yeah. So over the last five to 10 years or so, the rate at which literature is published has just gone up exponentially. I don’t remember what the last figure was, but it’s like tens of thousands of research articles published every single day in high quality peer reviewed scientific journals. It is no longer possible to actually keep up in the old school way. You can still get article alerts or whatever for particular topics of interest, but most people don’t have time to dedicate to that every single day. A lot of times things are sort of like just in time education, like, “Oh shoot, like I haven’t taken care of a patient with this particular issue or disorder for some time.” And I need to make sure that I’m up to date because, even if it’s just been a few months, you could have missed some major developments.
Generally there’s two ways that I use it. Like, I use it in my spare time or, let’s say I’m working on a project or background or whatever planning to give a talk. Then, even on something that I know quite a lot about, because so much changes so quickly, I will use AI, one of the tools. I’ve tried multiple of them and we can talk about that in a minute. So I use it in that way. I’ll ask it to pull all the relevant papers on a given topic and I specify the kinds of journals that I want them to come from.
And then there’s the sort of just in time learning where, like, I’m on clinical inpatient service and I’m taking care of a patient with a rare complication or a rare disease. And if anybody tells you that they know everything about everything in their field, they are lying, or have very little insight, because it’s just not possible. So when I encounter those things, I’ll do a search. I don’t put in any identifiable information about any given patient, but I might give the general, “That’s a patient with this, of this age, with this disorder being treated this way, is having this complication. And I want to understand more about how to manage this particular complication in this context.” Right? And then it’ll give me papers.
In pediatric bone marrow transplant and cell therapy, so much of our data comes from adult literature. So the amount of actual, dedicated pediatric research for many of these issues, whether it’s drugs or complications or other aspects of care, they’re largely derived from research done on adults. And you might get a handful of articles that are pediatric specific. So what I’ll do is obviously I’ll first see what’s there and I of course also validate the articles that it has told me exists. I always ask for links and I check them out. I don’t just assume that it’s accurate or what it says it is. And then, depending on whether I’ve seen the papers before, I might quickly skim them just to see if there is anything really new that’s developed, or this is just somebody else who’s reporting some similar observation or different aspect of something that’s sort of already known, nonetheless important.
I don’t need AI to summarize that for me. But when there’s been a lot of developments or I’m trying to catch myself up on, on something that I haven’t necessarily thought about for an extended period of time because there’s just so many things that happen, that go wrong, or decisions that need to be made because there’s a unique circumstance for a given patient. So if there’s a lot of papers that show up that have different approaches or new information to me, then I’ll have the tool give me a synthesis across all the articles that have been returned.
And then, based on my personal detector that I’ve developed for what is likely to be inaccurate, what do I need to make sure is not a hallucination, then I gauge how deeply I need to. You know, if I’m using that summary for some sort of writing or talk or public, then there’s no trust. It’s all verify.
Karen: So do you ever use it in addition to a research assistant? Do you ever use it to help you with writing or with generating?
Joanna: I do.
Karen: Okay.
Joanna: Yep. That’s actually where I get a lot of benefit because I can be verbose. I think that might be evident in even my speaking style. And so many times there are word limits, whether it’s crafting a resume that needs to have nice tight bullets or whether I need to write an abstract that’s no more than 250 words based on this entire research paper that I’ve written. My strength is not brevity, and I know this. And I could sit there, and I’ve had to, up until the last couple of years, when these tools were not readily available. I would sit there for sometimes days trying to cut words out of something. And, you know, you get married to sentences, you’re like, “Oh man, I really like the way that sounds.” And you don’t want to get rid of it! Anyways, that, I find AI is excellent at it. And you can iterate on that endlessly. So if I’m like, “Hmm, I don’t really like the way that sounds”, you just click again. It spits out something else. And I’m like, “Oh, okay. I like that better.” I may do it a sentence at a time, depending on whether it’s a particular sentence that I’m like, “Oh, this is a run-on, and I’ve got to figure out how to make it less.” I don’t know how to write for scientific journals. Then I’ll pop it in. They’ll come up with something that is on point, in a fraction of the time that it would’ve taken me to try and go through and get unattached to what I’ve already written and rewrite it in a better way. So I think that that, for me, is probably the way that it’s been the most valuable.
I also will sometimes use it to create templates for myself. Particularly for patient care, because even in delivering care, getting the information out in an efficient way is a challenge. Every system is different. I’ve worked in health systems with different electronic health records. And each deployment allows you to do different things, different than others. And so even just being able to get the information out in an efficient way for when I’m trying to round on patients in the morning, that is more difficult than it should be. So sometimes I’ll use Doximity, for instance. That actually I found to be incredibly useful for that particular use case. Obviously it’s not something that non-physicians would be using or have access to, but there’s other tools. ChatGPT does it too. Gemini did. They all do it. They can all make you a template for something, which I find also to be very useful. I don’t have to then obsess too much about exactly how I want the information to be organized. A lot of times it’ll generate something. I’m like, “Oh, this is so smart”. I might have come to that eventually, but this took a fraction of the time.
Karen: So do you ever use AI tools in your home life? Or are your kids being exposed to AI tools in school now?
Joanna: So home life, well, I use it to make our family calendar in Canva. Mostly because I get a little bit of my artistic itch out by playing with some of the design tools and whatever. So I do use that.
My kids go to a STEM Charter school. So they have unique technology-forward education. They’re not equivalent to unlimited screen time, but they actually learn about how to use some of these tools. And they learn about critical thinking and how to analyze that answer that something gives you, or what’s a reliable source of information, those kinds of things. It’s certainly not unsupervised, I will tell you that much. I’m not sure that there’s any tool yet that can deliver much more than a rewording, something that you’ve already created, or a template that you’ve asked for that doesn’t require significant checking to ensure that what it’s given you is accurate.
Karen: It sounds like you’re making good use of AI tools in a lot of different ways. I’m wondering about when you choose not to use AI tools specifically. And also, as a physician, what are your thoughts about people that ask “Dr. ChatGPT” for advice, like they ask “Dr. Google” for advice?
Joanna: I definitely don’t use it to guide my medical decision-making. So even if I use it to pull me a synthesis across journal articles, first I verify that the journal articles are ones that are high-quality, published and peer-reviewed top-tier journals. Even if that’s the case, that still doesn’t guarantee that it’s a well-designed study. But that would be a digression. Any time I’m using actual primary research papers to guide my medical decision-making, I’m always making sure that the study design and analysis apply to my particular population and question that I’m trying to answer.
The analogous situation is for lay people who are asking “Dr. ChatGPT”. So it may feel like you’re receiving authoritative information that is accurate and maybe has some literature attached to it. Even if you ask for links and citations – until I really did my Master’s of Science and clinical research, even I didn’t realize what it takes to design a high-quality study that’s applicable to a population of interest to answer a given question. And, you know, we joke in my epidemiology and biostats classes that there’s only better and worse ways to do things. There’s no perfect way to design a study. So you can find a problem. You can find an issue with any study design. It’s understanding what’s relevant and what’s not relevant.
And the same thing applies when people are asking. My family, for instance, they’ll tell me about a symptom that they’re having, and they’ll tell me what they think is going on. And I’ll ask them some more questions and they don’t even understand why what I’m asking is relevant. And then they’re able to say, “Well, why are you asking me this?” Then I can say, “Well, because X, Y, and Z and this is why it’s important in this context.” ChatGPT, or whatever you’re using to ask about your own medical condition, is not going to come back to you and say, “Well, how about are you experiencing this or that? Or what about this aspect? What other medications are you taking?” It’s going to answer the question that you ask. But the issue lies in the things that are not offered by a patient. And this is true even in patient care. The answers are a lot of times in what’s not said, that you have to ask about specifically. And that’s one of the things that we teach trainees, during their residencies, after medical school.
And there’s also this old joke, right? The intern will go ask a patient a question and we’ll give ‘em one answer. Then the resident will go ask and they’ll give ‘em a different answer. And then the attending will ask and they’ll give ‘em a different answer, you know? And the attending gets the most correct answer very often. But what you learn is, the reason is, it’s not just because people don’t remember. It’s that asking people in different ways generates different answers, much like ChatGPT does, right? When you ask a question in a different way, you get a different answer.
So I have in the back of my mind, because of my medical training and years of experience, what are the things that I need to make sure of? What are the serious things that I need to rule out or make sure that we identify urgently if there are things that need to be dealt with urgently? What are the can’t-miss things? And then, what’s going to contribute to actually figuring out what the problem is? And both of those things are important.
So, the issue is not about what gets asked, it’s about what doesn’t get asked or offered. I have so many experiences where I’ve had to ask the same question. In many different ways. And sometimes people will be like, “Why do you keep asking me this?” And they might get a little frustrated and I’ll say, “Well, here’s what I’m trying to understand.” I’m not being deliberately confusing. It’s just that sometimes I’m not quite getting the information that I’m looking for. And so sometimes just asking in a different way will help somebody understand what it is that I’m trying to get at.
And it’s those nuances that are lost in this technology at this point. And I’m not sure that it will ever be able to do that. Because again, it’s just like a problem of documentation. If it’s not there, then it can’t be trained on it. And so much of what happens to people happens outside of the hospital, outside of the clinic. So even using medical records, that’s a snapshot in time. It’s no different than doing an EKG to check somebody’s heart cardiac rhythm. That’s just, in that moment, there’s no arrhythmia. That doesn’t mean that an hour from now you’re not going to have a heart attack or have some other abnormal arrhythmia. So it’s about understanding the bigger picture and that’s all the intangibles that are not part of the chart. Even if you were to use a scribe, there’s so many intangibles that go into medical decision-making.
I’m sure there will be some people who are, like, “Whoa, that makes me nervous”, you know? But honestly, that’s the whole reason why it’s called the practice of medicine. If you could just use a textbook, we wouldn’t need doctors, right? Everybody could just grab a textbook and figure it out. The thing is that, in training, in pediatrics in particular, during your intern year, the goal is by the end of the year for you to be able to, like, from the doorway, it’s your “doorway exam”. So I open the door. Is this patient sick or not sick? And that’s the goal, to be able to see. And that’s like a gestalt, right? That’s just reading the energy in the room, and how does the patient look, and all these little factors about how they’re behaving and what you know about what’s gone on. And your instant observations, which are not translatable to an algorithm.
Karen: When we talk about the AI tools that you use, do you feel like the providers have been transparent about where they got their data? And I’d like to recap how you got the data that you use for your study.
Joanna: Oh, yeah. So the data that we use for our study, this was reviewed by the institutional review board [IRB]. There was a regular protocol that was submitted that is consistent with the regulatory requirements for using personally identifiable health information in research. And it was all de-identified. So when you use it, it’s de-identified, but obviously you have to access the charts. So the research team is going to have some knowledge of personally identifiable patient information. But you have protections in place, with your data safety and monitoring plans and data protection plans and protocols for the data that’s allowed to be collected from what timeframe. It’s not open season to collect data for the rest of eternity. That all went through all the same processes and protocols that would’ve been required of a drug trial or a device trial or any other research that was involving patients or patient information.
So we knew what our population was. We had our timeframe. We partnered with the business intelligence folks who can help identify the patients. But there could be a whole separate conversation about how I actually identify populations of interest in the EMR longitudinally, across time, and follow individual patients longitudinally. That’s another animal. Once the patients were identified, then we did data mapping. I sat there with the computer scientists and we looked at the various places that this data lives in these records and identified those places. And that’s where it got pulled from. And sometimes it was multiple places. Then there were technical ways of ensuring that data was selected from a primary place or whether it needed to be a composite data point or whatever the case may be. But we had access to the information only because we submitted a research protocol to the Institutional review board, which was reviewed and approved. And nothing, no data was ever accessed prior to having that written approval from the IRB.
Karen: I’m curious, did the protocol call out any representation across, say, gender or race or any other aspects where the data could have some biases that might end up being reflected in the study results?
Joanna: Actually, I think it’s really important to include those personal demographic information about patients. Because as you said, there are socioeconomic determinants of disease. There’s definitely plenty of data about disparities in care delivery and access to care, based on race and ethnic and or ethnicity. And if you don’t collect that information and look to see if you’ve got representation in your population, then you can’t say that it’s generalizable across all people. It goes back again to making sure that you are applying whatever tool, algorithm, data collection strategy, whatever your goal is with your research that, and long term who it’s supposed to apply to, that you’ve got a sample included in your trial that is consistent with what is in the population that will ultimately, hopefully benefit from what it is that you’re developing.
So that’s an important thing to include, to ensure that we have a full understanding of how it impacts all the people, not just, you know, white people.
Karen: Right! So as consumers and members of the public, and people who use these systems, our personal data and content has probably been used by some AI based tool or system. I’m wondering if you know of any cases that you could share either where your data was obtained properly or maybe obtained improperly.
Joanna: Personally I’ve gotten, over the last couple of years, several letters from providers from health systems that I’ve had care at, talking about data breaches. And I’m aware only because I actually looked to see how these things are often used, because I was unsure of exactly what people like criminals would do when they breach a health system, other than trying to assume somebody’s identity for purposes of fraud. But it turns out, people can not just assume their identity, but they can actually use that identity to get medical care and use insurance benefits and things like that. So it’s not just about having your Social Security number and stealing your credit card information somehow, or another opening credit card in your name. It’s about actually showing up as somebody else and getting medical care that then ends up in your medical record.
And these are all things that I read about last night after you had mentioned this question. And it’s wild because I would never have – I mean, I read about a patient who received a kidney transplant under the guise of being somebody else.
Karen: Wow.
Joanna: Yeah. I think the BBC had written about this. And I know that Anthropic has publicly addressed some hacking issues. I believe that those issues did impact the use of medical information. To my knowledge, nobody’s impersonating me. But I do know that my information was at risk, or was known to be stolen for some purpose.
Karen: Yeah, I just got one of those breach notices myself last week. And it’s not the first time that I’ve gotten one, right? It definitely happens. But I really have to wonder. It’s not an excuse for fraud, but if people are uninsured and they’re desperate and they need some way to get care, this is the only option they feel that they have – you know, that’s pretty sad.
Joanna: I totally agree. Yeah, that’s a much bigger problem.
Karen: You know, the hackers that are stealing their information and selling it to those people, they’re just further exploiting the people.
Joanna: Right, exactly.
Karen: That’s not good either.
Joanna: Totally. Yeah. That’s certainly not the solution to the problem.
Karen: No.
Joanna: But yeah, it’s interesting because as I was reading about this, it sounds like “vibe hacking” is what they call it, which you may be even more familiar with than I am, because I was just learning about vibe hacking last night. So you can just tell it what you’d like to hack into and it’ll write you some code to try and breach. I’m sure it’s not quite that simple, but it seems like it’s pretty straightforward. As far as I can tell, there’s not that many people who can successfully hack a system. But maybe that’s just me being naive.
Karen: Yeah. Yeah. As far as your personal information, do you take precautions as far as the information that you share, like on social media sites and such?
Joanna: I only put information out there that I don’t care who has it, because otherwise, all bets are off. So I don’t write anything on social media that I might regret in 10 years. I don’t post personal information. I don’t have my location attached to the pictures that I post. I try and remove any metadata from anything that I might share. And that goes for actually when I’m sharing PDFs of my resume, for instance. I remove all the metadata.
But honestly anybody who’s Googled themself before, as I have, you can find out easily — like, who my sister is and where she lives, and who my ex-boyfriend was from 10 years ago. It’s amazing the things that you can find about yourself online. So honestly, I only put things out that I don’t think could be harmful in any way.
But as far as telling Facebook that they can’t use my data, or LinkedIn that they can’t use my data, I mean, the data that they’d have access to is out there from my standpoint, because that’s sort of how I’ve tried to protect myself. And honestly, if Facebook wants to show me ads for clothes that I might like on Amazon, I can spend some time scrolling and see some things that I like.
Karen: So you mostly live by what they call the “billboard rule”, which is that you don’t put anything online that you wouldn’t want other people to see on a billboard?
Joanna: Totally. On a billboard. Yep. That’s exactly right. Otherwise, most things can be found out about people whether you try and scrub it from the internet or not. I mean, I’ve got antiviral, I’ve got McAfee that has the ability to pull things from the internet and scrub it from time. I’m sure it does a reasonable job, better than having nothing, but it’s not like things are not going to just pop right back up in the next day or two.
So if somebody really wants to get at your specific information, I feel as if they’re going to be able to do that. I don’t know that there’s anybody who can avoid that. Maybe if you’ve got ultimate security around your person, your assets, but even then, I don’t think anybody’s immune to that.
Karen: So has any company’s use of your personal data created any specific issues for you, like privacy or phishing or loss of income?
Joanna: Yeah. I’ve been doing some consulting. I’m also open to other regular employment opportunities within biotech and pharma, related to clinical development, both early stage and late stage. So I’ve been engaging with a lot of recruiters on LinkedIn. There’s plenty of people who are scammers. Who knew that? Why would somebody impersonate a recruiter? I think there have been two times where I fell for it. And the first time I was chalking that up to naivete. The second time, that person was really good and it wasn’t until I stepped back and I was like, “Wait a second. They’re writing a whole lot more in this email than most of the recruiters that I’ve encountered.” And that was after several emails and then asking one of my friends (who’s a recruiter) what her opinion was, like, is this legit? Because now that I’m looking at this job description they provided me, it looks an awful lot like what my profile says. It’s kind of perfectly suited and says some of the same words that I say about myself.
So far, I’ve only lost time and I’ve had some pain and suffering from annoyance and frustration that I actually fell for it. But thankfully I didn’t part with any money. But I can easily see, there’s a universe even where I might make that mistake. Because I just think that people are really good at impersonating somebody when they’ve got a target. Particularly with the job market the way it is, you hear from a recruiter who’s contacting you about a job opportunity, most people are like, “Oh my God, thank God.” So it’s easy to do. And the first time, my clue after I had gone back and forth and wasted several hours with this person, they asked me to switch off the platform and move from LinkedIn to Fiverr. And then I was like, “Okay, alright. You might as well have asked me to go to WhatsApp.” You know?
These phishing attempts, there’s plenty of other ones. Those are the ones that I fell for at least enough to engage with them and realized that I just wasted a huge amount of time. I get plenty of other ones where it’s obvious to me. They’ll send me direct messages on LinkedIn, posing as recruiters. I just report them to LinkedIn every time.
I’ve gotten messages from people on Facebook. I’ve gotten emails. I don’t know how people have gotten my Gmail address. I mean, granted, I’m sure I’ve used my Gmail in enough things that it’s not too hard to find that, right, but when you get a personalized email from somebody that you think you know, as a pediatric oncologist, my default mode is trust. I’ve had to work hard at trying to be a little more cautious over the years, for my own safety. And not everybody has good intentions. And not everybody is being honest about who they are and their reason for contacting you.
Karen: Yeah, learning how to be cautious, it’s kind of sad that we have to, but I think that’s just a reality of life nowadays. So I really appreciate all the time that you’ve shared with me for this interview. Is there anything else that you’d like to share with our audience? Any final closing thoughts?
Joanna: I guess just: I would never discourage anybody from educating themselves. Please, by all means, read about your symptoms. I don’t mind one bit when my patients bring me articles or ask me or say, “I’ve read about this, and what do you think? What are your thoughts on this?” I welcome that. As do, I would say most doctors, and certainly all the people that I’ve ever practiced with. So I don’t think that patients should come away from this or other similar conversations, hearing that we don’t want them to be empowered or be in the dark. We want to be able to help guide them toward useful information that’s applicable to them and that is going to actually help whatever it is that we’re seeing them for. Not all sources are created equal. And there’s lots of misinformation out there. There always has been and just is getting worse. So I think that I would encourage people to feel empowered to bring that information to their doctors and ask directly. And if their doctor says, “Hmm, I don’t really think this is a great source of information”, then answer back and say, “Can you give me something that’s better? Where should I be looking for this information? Or do you have anything that you can offer to me?” I will often pull articles for people or give them a list of websites that are freely accessible where they can get high quality, evidence-based information that is reputable.
So I think that that’s probably the thing that I want people to come away with is not to construe when doctors are discouraging people from using these tools. It’s not because we want to be in control, it’s because we really do care about making sure that people are getting appropriate, timely care that is in their best interest.
Nobody becomes a doctor because they want to sacrifice their twenties, thirties, and get yelled at all the time, if they didn’t really care about making sure that they could help people feel better from whatever it is that ails them.
And I think with that in mind, as you know, we welcome questions and additional explanations. I at least always tell my patients they can ask me the same question until they get an answer that makes sense to them. And I think that that’s really important. There’s no silly question. There’s no question that they shouldn’t feel like they can ask their physician to help with. There shouldn’t be a need to feel like you have to use ChatGPT, otherwise you can’t ask a question. I guess that would be a bigger problem, could be an issue between them and that one person. But I think it’s just important that people understand that we want them to be knowledgeable and empowered, but with the right knowledge, because that’s what’s going to help them get better at the end of the day. We can’t do everything. We can’t make them take their medicines. We can’t eat right and exercise for them. But we can sure help guide them toward the resources that are going to support that.
Karen: That’s a great summary. Thank you so much Joanna, and thanks again for joining me on this interview. It’s been a lot of fun and really interesting to learn about everything that goes on behind the scenes, so thank you.
Joanna: No, I appreciate it. This was great. It was fun.
Interview References and Links
Nimble Healthcare Innovations & Services, LLC
Joanna K. Grossman, MD, MSc on LinkedIn
Joanna Grossman, MD, MSc on Substack
About this interview series and newsletter
This post is part of our AI6P interview series on “AI, Software, and Wetware”. It showcases how real people around the world are using their wetware (brains and human intelligence) with AI-based software tools, or are being affected by AI.
And we’re all being affected by AI nowadays in our daily lives, perhaps more than we realize. For some examples, see post “But I Don’t Use AI”:
We want to hear from a diverse pool of people worldwide in a variety of roles. (No technical experience with AI is required.) If you’re interested in being a featured interview guest, anonymous or with credit, please check out our guest FAQ and get in touch!
6 'P's in AI Pods (AI6P) is a 100% human-authored, 100% reader-supported publication. (No ads, no affiliate links, no paywalls on new posts). All new posts are FREE to read and listen to. To automatically receive new AI6P posts and support our work, consider becoming a subscriber:
Series Credits and References
Disclaimer: This content is for informational purposes only and does not and should not be considered professional advice. Information is believed to be current at the time of publication but may become outdated. Please verify details before relying on it. All works, downloads, and services provided through 6 'P's in AI Pods (AI6P) publication are subject to the Publisher Terms available here. By using this content you agree to the Publisher Terms.
Audio Sound Effect from Pixabay
Microphone photo by Michal Czyz on Unsplash (contact Michal Czyz on LinkedIn)
Credit to CIPRI (Cultural Intellectual Property Rights Initiative®) for their “3Cs' Rule: Consent. Credit. Compensation©.”
Credit to Beth Spencer for the “Created With Human Intelligence” badge we use to reflect our commitment that content in these interviews will be human-created:
If you enjoyed this interview, my guest and I would love to have your support via a heart, share, restack, or Note! (One-time tips or voluntary donations via paid subscription are always welcome and appreciated, too 😊)
