AI and Healthcare Virtual Summit. I'm presenting my experience with mining the medical record for bone health. By way of introduction, I'm an osteoporosis physician. So I'm a practicing clinician, but I'm also the executive director of Maritalupe Diari-Gamal, which is the Sydney Partnership for Health Education Research and Enterprise, otherwise the acronym is SPHERE. And the areas I'm going to cover in mining the medical record is basically what is research translation? And a research translation center in Australia is very much operational. It's about how to get the evidence that no doubt you've heard a lot about and applying to AI in health, how to get it into practice. And my own experience or near death experience of moving from a very rich mining source of a standalone clinical record for clinical trials and how that morphed into fractal IOs and almost dragged towards fractal IOs by technology enabled enhancements. And then stepping back from the mineshaft and taking a high level view are the things that we can actually learn from large scale mining operations of anything, whether it's dirt, you know, minerals or data. Some of the principles are actually very similar. So just by way of introduction, what is SPHERE? These are Australian NHMRC, so government, federal government accredited research translation centers that constitute a relationship between health service providers and universities where in Australia, the universities by and large aren't providing healthcare, they're doing health and research in healthcare, but aren't the healthcare providers for the large proportions of public health patients. And also seven medical research institutes, which are more vertical in the integration of research interest in genomics, dementia, cancer, for instance. This relationship actually covers the facility-based care for more than 2 million people in metropolitan Sydney. We have more than 60,000 employees and $8 billion in combined revenue. And as alluded to at the start, the mission is actually delivery, it's implementation rather than discovery. The discovery happens in those university and medical research institutes to some extent in the health service providers. And there's a vehicle required to get discovery into practice and bring innovations to life. And so the three things you actually need for that to happen is actually opportunity, talent and desire. And in implementation, those are three non-intersecting circles and the magic happens when you can actually get those three things to intersect. So in digital health delivery, you must have executive support providing the opportunity and without that executive oversight and approval, then nothing really is going to go forward. You must have the clinical champion. In my case, as an osteoporosis physician, I was the champion for the bone health aspect of this presentation that I will present to you, but you cannot have just one champion alone. Often you need a team of across sites and also intersecting into the community so that you get that leverage from the community saying that this is what we need and want to see delivered. And the third is how do you actually manage change? And we'll touch on that perhaps at the end, but I can preface it by saying that you need to provide for change to actually stick. The incentive for change is proportional to the likelihood of sustainability. Or if you're going to bring your workforce with you, then the change to the, because we are creatures of habit, any workflow change needs to be minimal or is almost inversely proportional to the likelihood of sustainability. And the magic happens where those intersect. And you can imagine there is some noise around that and getting them to intersect is actually the art of this job. So what is the problem we're trying to solve? Well, the problem we're trying to solve is the same problem I had 40 years ago, which is preventing osteoporotic fractures, either primary prevention, which is stopping them from occurring at all, or secondary prevention, which is making the first fracture the last. And during 40 years of clinical and practice, we have actually gone, I will go out on a limb and say that osteoporosis is a curable disease, not just treatable. It's been treatable for decades, but it's a curable disease where we've moved from understanding the relationship with age and menopause and the effect of hormone replacement therapies on fractures, which is certainly where I started. The introduction of bisphosphonates, which revolutionized our ability to actually treat osteoporosis, anabolic therapies that became more prevalent in the 2000s, and then the explosion in biologicals for both anabolism and anti-catabolism, which has revolutionized our approach to osteoporosis. So we have largely, I believe, cured osteoporosis, and the problem now is getting those discoveries into practice. Simultaneous to each of those has been data-driven milestones where we have improved our ability to actually interrogate, acquire, and understand data, not least starting in the 1980s with the development of DEXA as a tool for measuring osteoporosis via bone mineral density. The onset of understanding the interaction between bone mineral density and fracture risk calculators for osteoporotic risk calculators in the 1990s and their widespread use. Introduction of electronic, or not introduction, but widespread scale-out of electronic health records in the 2000s. Better understanding of machine learning and introduction of machine learning in the 2010s, and harnessing the creative opportunity presented by COVID as we went to telemedicine and look at the way we're doing our virtual summits these days. So becoming more comfortable with both telemedicine and digital health, and now the revolution of generative AI and what that will allow us to harness. And so my journey from a clinical trials-based unit to a fracture liaison was almost crisis-driven and you should never let a good crisis go to waste as many have said over the decades. And this was actually my crisis. For two decades, I had been mining my own business, not minding my own business, mining it, which was we had the densitometers from numerous sites and numerous machines feeding into a standalone database that captured all of their clinical information, which was required at the time that the bone density was being captured, that could inform fracture risk, which is now actually standard in most of your machines, but we've been starting this from the early 2000s. And so we had a large standalone database with about 30,000 patients that you could actually interrogate for randomized control clinical trial recruitment to any number of studies that were underway at that time. And you could identify on age, gender, you'd just query who was available and what therapies they were on and then make your recruitment targets. And so that was a very neat, tidy and functional data environment until in 2010, 2011, there was a recruitment request for what was our first head-to-head comparison of a Lendronate with Rheumatozumab. And the recruitment criteria for that posed a problem for the standalone dataset that I had, which was that while we had numerous post-menopausal women in the age range with the bone density as indicated on this slide, what we hadn't captured was the severity of their vertebral fracture or how many they actually had, because you only needed to know that they had a fracture for bone density to qualify. And so we didn't have the severity or the time signal. So for recruitment into this study, there had to be a proximal femur had to be sustained somewhere between three and 24 months. And so what I did in response to this crisis was turn around to the hospital mainframe because the database was standalone and came to the hospital mainframe, which had the coding data for patients who were admitted with any of these outcomes with vertebral crush fractures or with hip fractures and the time intervals. And I did a calculation of how many were admitted and made an assessment of how many of those I'd be able to capture. And of course the crisis was I under-recruited compared to what I thought was being admitted. And so that led to in a breakdown of the failure to meet our recruiting numbers, what had actually gone wrong. And it turned out that others were struggling with the same problem, probably for different reasons, but with the same problem. And in Sydney, we had four investigators who were looking at AI and coding mechanisms for identifying patients at risk of osteoporotic fracture. And we were all within a 10 kilometre radius of each other. So we collaborated and Rory Clifton-Bly at North Shore, Royal North Shore, had developed a coding-based and a text-based searching tool for patients with osteoporosis that he had called AES. And I had worked with my colleagues in industry and academia to develop a natural language processing model, which I called X-ray. And that tool was actually reading the radiology reports in real time, because the reason I had not been able to recruit was I hadn't been informed of the patient at the time of their admission to be able to recruit them into the study. And retrospective review of coding meant that I was late in getting to the patient. So, and so the tool that I'd used for coding to make my estimate of being able to identify patients with fracture was too retrospective to the event to be able to actually recruit on time. So now Marcus Seibel at Concord was curious to know if these things were identifying the same patient or different patients. And so I funded a study where at two hospitals, one Royal North Shore and the other at Concord, where over a six month period, they extracted the radiology reports for patients of both sites, just to see whether they would find the same patient or the extent to which they didn't find the same patient. And I remember at the study startup, we anticipated we would have exactly this kind of overlap. We just didn't know by how much. And so that data is published and I want to acknowledge Kendrick Blacker and A.A.N.T.I., which you all know who did, you can imagine they did the lion's share of the shuffling through those reports and coding and checking that the tools had high fidelity, which they did. So while there was a degree of overlap between a coding and a language model for identifying patients with fracture, from the academic imperative was met, which we could identify how accurate they were at identifying patients with osteoporosis. From a functional perspective, we now had thousands of patients, we were finding thousands of patients that we previously didn't actually know existed. And that posed actually a delivery problem because once you've actually identified the patient with fracture and you've actually got treatments and now cures for the disease, you have a medical legal responsibility to be able to respond to it. And so in many ways, you have to be careful what you wish for because my clinical practice and clinical trials research organisation very quickly had to redirect its resources towards fracture liaison services. And we actually knew, and at the same time, Nick Pocock, who was also within that 10 kilometre radius, was working with computer aided diagnosis of vertebral crush fractures. And we knew from the study of osteoporotic fractures in the 1990s, and this is data from Dennis Black, that the presence of vertebral deformity strongly predicted the likelihood of both further vertebral deformities and hip fracture, which imposes a massive cost on the health services for their repair and management. And in multivariable models, the relationship was extremely strong, both for the prevalence of vertebral deformity, both in terms of the number of vertebral deformities that you actually had, as well as their severity with the greater the severity of the crush fractures and the higher the number of vertebral crush fractures leading to extremely high levels of risk for both vertebral fracture and hip fracture and non-vertebral fractures. So we already knew that that relationship existed. And the question was, do computer aided diagnosis of VCF, of vertebral crush fractures, are they clinically useful? And so Nick did a study again over a five month period, comparing a validated vertebral fracture identification tool using AI of the CT images of thoracic and abdominal CTs, coincidentally identifying vertebral crush fractures. And while his conclusion from an academic and perspective was that the sensitivity and specificities were as reported and that the negative predictive value was quite high, it probably couldn't be used clinically on the basis of its positive predictive value. But from my operational perspective, the drama was actually in table two. And I'll just take you through that. For those who aren't familiar with severity scores, there are four grades of gonad severity for vertebral crush fractures with no vertebral deformity being grade zero, grade one being minor. And by the time you've got grade two and grade three deformities on lateral spine x-ray, the likelihood of it being osteoporotic is extremely high. And so we just concentrate on those with the highest likelihood of actually having osteoporotic fracture. Table two showed that of the 183 who were shown as true positives, 62% of them had had that fracture reported, routinely reported in the report that had been provided with the original scan. And the computer-aided diagnosis had found another 72, or there was a 65% increase in the number of vertebral crush fractures that were being identified by this. And that the number that you needed to screen was about 20 or 23 that were required to be screened before you'd find a new fracture. That would pose an incredible burden on the fractal liaison services. And this is where the concept of actually mining the medical record occurred to me, because we'd actually started at a very high level of review, which was actually at the densitometry level. And then coming in at coding and then moving down to the reports. But if you then move down further into the digital record, you are just finding more and more and more patients undiagnosed with osteoporosis for whom the ethical problem is you now have documented treatments that work and potential cures for some. So you need to be careful what you wish for with technology because it may actually come true. And the outcome of this story was that to my relief and probably not to the benefit of the patients, but almost to my relief, there was a system upgrade of the radiology reporting system, which actually was incompatible with the innovation and the natural language processing that I had actually installed. And so the administration, the executive turned off the AI by almost by default, because with the system upgrade, the AI didn't come with it. But I had had a look over the edge and into the mind pit. And the biggest problem that we've got is actually the osteoporosis treatment gap. And these are meta-analyses recently published by Ali and others, a meta-analysis of the reported fractalized services. And the take home message is that not all eligible patients are being treated with about best in this meta-analysis, 77% being treated. But even those that are being treated only, you know, the majority are, well, I think about 70, I've just got my own 72% of patients who are dispensed, who are given a script, only 72% actually dispensing. And furthermore, if you, next slide, hang on a second. I've stopped the move forward. Yeah, this, and on the next slide, the only minority of the patients who actually start are still persisting at, at about three years, with about anywhere between 20 to 40%, who are still on therapy at three years. So we actually, that's called the osteoporosis treatment gap and finding more patients. You need, you need a whole system of view, view. You need to understand the entire clinical interface, or you're going to find that the, you will be finding more patients for your fractalized and service than can actually be treated or will adhere or change the outcome of the disease. And while fractalized and services have different characteristics in different, in, in different contexts, we follow here. And I'm certainly familiar with Marcus's model here in, in, in Sydney, published by Curtin Gander in 2013, which looks like this. And we, we followed the, we were trying to follow the type A FLS model, which he had shown was the best for adherence to, to osteoporosis, to effective osteoporosis therapies, which was to identify, educate, evaluate, and start the treatment and continue that treatment and follow up at 12 months for, to confirm that there is adherence. The type B model where you actually identify, educate, evaluate, and then defer to the primary care physician for treatment is also an acceptable model for the others that are essentially not going to work. Well, that was published in 2013. And with AI, and what's happening now is that the computer aided diagnosis is identifying vertebral crush fractures. And there's almost a type E FLS model, which is the radiologist is identifying, and there's no education, and they just defer it to the primary care physician and hope for the best. So that's a significant gap and AI isn't contributing to the isn't solving the problem. In fact, it's amplifying it. And so in understanding the osteoporosis treatment gap, there are a lot of stakeholders and a lot of variables with both, as I indicated, the willingness of the institution to actually provide the infrastructure and resources to take on the responsibility for fractional liaison for both its identification of the patient and for their treatment. And the you know, who is actually in the third party payment of the therapies and the role of the primary care physician, I think there's actually an opportunity to actually look at the patient and the complexity and the medical health record to be able to identify the patient who would be at greatest benefit. And this is data from Greenspan's group, which looked at the health economics of treating identifying and responding to patients with osteoporosis. And the greatest potential is actually in those who are older, but at the lowest risk of dying. And so while there is a unequivocal cost benefit to identifying patients with insecure fracture prevention, and treatment over observation, these, the incremental cost effectiveness is greatly improved, which is a, you know, argument that I can take to the executive, if they are older at initiation with least likelihood of dying after fracture and potentially less comorbid diseases that contribute to that competing risk of death. Also, the likelihood of actually having the most expensive fracture, which is hip fracture, will also strongly influence the incremental cost effective ratio. And so that opens the possibility that I can integrate the presence of fracture, and the the likelihood of a second fracture with their comorbid risk factors, including not just age, but comorbidities and impact on on the competing risk of death, to identify the patient who would be most important, you can actually create a priority list for those for whom your restricted resources in a fractionalised and service should be targeting. And so having had a near death experience from an AI enabled fracture identification system, I was wondering about what lessons we could learn from from mining in general. And this is a helicopter view or a satellite view. And they say, if you follow the yellow brick road, you end up in Oz. And this is Terra Australis, which is the largest island in the world and the smallest continent. And just by way of context to the to the to the audience, Australia is about 80% the size of China, about the same size as continental United States minus Alaska. And prior to the American War of Independence, was was, was only colonised after the American War of Independence, when convicts were no longer sent to Virginia. That raises a really important point, which is that just because data and land is available, it cannot be assumed to be used and usable, because we have 60,000 years of custodianship in Australia that that needs to be acknowledged. And at this point, it's a tradition in Australia to always acknowledge the traditional custodians of the land. And I'm talking to you from from Gadigal country of the Eora nation, and that data and land is never ceded without consent. And the name of Maradaloo Bidyari Kamal is a gifted name for the Sydney Partnership of Health, Education, Research and Enterprise. And it was generously gifted by the Aboriginal people of of the Sydney Basin means coming together for good health and well being. And so when we are mining data, all that the AI revolution has done is to actually improve the data extraction methods and processing with advanced analytics, which with our industry and academic partners, allows us the great capacity, as I've indicated, and experienced to identify more patients with fracture, but we actually need to sit back and actually look at the at the integrity of the data source. And this is the data source in which I'm operating, which is in New South Wales, it's about New South Wales and Victoria probably combines about 80% of the Australian population is about the size of Alaska. We actually are moving towards a single digital patient record of 8 million people across the entire New South Wales environment. So for hospital based care, which is currently 228 hospitals providing 300,000 surgeries, of which 6000 in any given year is hip fracture repair from osteoporotic fractures. And Australia why that's getting up to around 4 billion Australian dollars, so 2 billion US dollars a year in in a treatable cost. And those local health districts, 17 local health districts for four of those 17. They're my bosses, they tell me what they want to prioritise. And getting osteoporosis onto their radar is a key outcome. Most of you, however, will be more familiar with data environments that look like this, which is incredibly fragmented. And Australians will recognise this as the as the tailings of the Cooper PD, Opal mines. And basically what that's trying to represent is that data is often siloed and fragmented across numerous primary care, aged care, mental health, disability, pharmaceutical, as well as academic institution with universities often siloed from each other. And there's actually opportunity in that if you can actually federate the data across all of those, you actually create a great opportunity for resource and data mining, but requires a data sharing agreement and a commonality of purpose to achieve. So in primary care, which is where the osteoporosis treatment gap is largely resides, in the current scenario, you don't have a mother load. And so the majority of patients are dispersed across multiple practice locations, with fragmented referral and service pathways. And our Australian attempt to actually create a digital health record, to my mind, included complex implementation and workflow changes without enough community champions. And therefore, while it had the highest level of executive support, missed out on those two other key enablers, which meant that it hasn't gone according to plan. So that leaves us with secondary and tertiary care, we have better integration of the medical record at the hospital level, at that high end, high cost interventional place. And that lends itself to things like cancer care and transplantation where that's not being done in primary care, it's all done within one centre. And so you've actually got a structure that lends itself to data mining and effective interventions. But where you've actually got that crossover between primary care and community care, and you don't have that vertical integration, then problems exist between those who are responsible for the identification and those who are responsible for the management, being dissynchronous to each other and creating a workflow problem. The assistive technologies, however, as I have shown you, are incredibly powerful, both at the coding level with AI scribes, allowing coding in real time, when in past that was really done retrospectively. But often coding is for a second, the primary purpose for coding can often be for third party payment and reimbursement purposes, and may not always often doesn't give you the exact data you want regarding the presence or absence of fracture, for instance. But generative AI and large language models, as I've used for reading radiology reports in real time to identify patients with fractures so that they could participate in clinical trials at a time when they were still in the hospital is an incredibly powerful tool that will sometimes get you into trouble. As I indicated, deep learning of the radiology of the image is incredibly powerful for a vertebral fracture identification and new methods of not even bone density identification, but fragility identification, which will either complement DEXA or completely replace it. And those studies that are ongoing, you might have already heard about, but their implementation in the entire health network will require judicious and careful implementation. There is also the prospect of personalized medicine in bone health analysis, that which we see with genomics, where we redefine fracture risk in the elderly for those at the lowest competing risk of death by integrating their medical record and their likelihood of longevity with the likelihood of fracture events. And lastly, pragmatic clinical trials will allow us to very more rapidly identify real world data post-implementation. The most obvious would be, is there a cardiovascular risk with the rollout of rama-sosumab and access to real world data in real time will revolutionize our ability to answer those important questions. The other things in data mining are the responsible and sustainable practice and research ethics and clinical ethics, they overlap, but they're not identical. And there are emerging challenges with regulatory compliance, both in data discovery and leaching of data out of various silos. That means that the technology, the AI, this is always the way, the AI is ahead of the ability of the organizations to actually accommodate them. And that really also includes priority populations and indigenous health. And that like land data is not seeded just because you have access to it. So that means we need to be very, very careful with safety and security, managing the digital risks. And I actually think that federated learning and federated systems are actually a, can, even though they are an obstacle to integration of data, they actually allow us to be more comfortable with managing or preventing data breaches or when data breaches occur, limiting the blast radius of any such event. There also needs to be effective cost management. For instance, the infrastructure issue that I encountered that a system upgrade meant that the AI that I had developed became redundant within one upgrade cycle requires a commitment from the executive at the highest level to maintain the infrastructure and the operation, and also any of the IP arising from those discoveries. You've got to bring the community with you. They need to have identical clinical and community alignment with what you're actually trying to achieve. And indeed actually advocate for it. And if you can have, if you have consumer advocates saying that this is a priority for the consumer, then you usually find that the political and the administrative wings swing in behind that. And there is, there needs to be corporate digital responsibility, ensuring health and equity and user content. And we need to maintain transparency with data, particularly for those in priority populations and, and for others for whom, from cultural diversity. And the last thing is actually managing the data lifecycle, very much like a mining cycle, particularly where we're actually sharing data with our academic and industry partners. There needs to be maintenance of data privacy, data integrity, and an end of mine life. And so in summary, mining the medical record for bone health has been an interesting and challenging educational opportunity. It it's incredibly powerful, but you've got to be careful what you wish for because it may actually give it to you. And we don't actually know how to use it all yet. I must have at the highest level executive permission and support with a business case and funding to proceed. The data surveys need to concentrate on that vertically integrated medical record. And if it's not vertically integrated, then we need federated systems that fill in the gap because patients do not just reside within the hospital system. For instance, they are in primary care and we need a whole of system solution for, for where they, where they are at various times in their health journeys. We need to workflow with synchronous diagnosis and treatment planning that is simultaneous to monitoring and payment processes. The second level is I must have clinical and community champions who advocate and lead change in health technology. Without them, you, you, you will not get implementation. And when you have both of them, you can have your community champions pushing the appropriate buttons so that the clinical champions are given permission to, to proceed and manage change to manage change effectively in digital health. Sure. Never waste a good crisis, but as it was my experience, be careful what you wish for. Nothing happens until everything happens very quickly. And sometimes you need to be careful with some of the full implications of what's actually going to occur, but you do need to bring the workforce with you. And the amount of change to their current practice will be inversely proportional to their sustainability. With that, I'd like to thank you for your attention and I'll take any questions. That was great, Dr. White. We're still waiting for the attendee to post some questions, because AI is new. And I just want to understand some of the things that you mentioned. So you start, what I understood is that you started by the AI reviewing spine imaging to identify people who have either non-stress fractures or can develop another fracture based on the previous fracture history, correct? We didn't start with imaging. We started with actually the radiology, actually we started with the coding, which is actually the medical record that had they actually been coded by the organization, which is the highest level of data often acquired for third party payment or classifications, right? So a colleague of mine had started with coding, right? And when I started looking for patients, we all go to the, have they been coded for having osteoporosis and approved unsatisfactory for recruitment in a clinical trial. So I went next level down into the, into the mine shaft, right? And the mine shaft in the medical record, the next level is actually has somebody reported it, but it didn't get coded, right? And so is it actually in the medical, is it in the radiology report? Has the radiologist seen it and reported it and it didn't get coded? And we found that that was actually the case. And then if you go the next level down, and so you can see how you're mining, you're going deeper and deeper into the medical record. The next level is actually the digital record, which hadn't even been reported. Does that make sense? And so as you went further and further down the mine shaft, you're finding more and more patients that hadn't been coded and hadn't been reported that actually had the disease. And so if you use AI for osteoporosis, you need to be able to respond to the, to what it is you're about to uncover because it's incredibly prevalent. Yes, I can see that. So we got a question from Nadine who is asking, what are the challenges you meet with mining on such a large scale? Yeah, the output, the output. And so it comes down to the, to the density. And so what I was trying to identify was that because I was mining a medical record that I, that I controlled originally over the last two decades, which was actually the bone densitometer, right? And so I used to mine the bone densitometer and your hit rate was very, very pure. So it was almost like gold mining and you're on a mother load. If you're actually mining the medical record and you've recruited because on the basis of bone density, then the, the, the richness of the data is very relevant to osteoporosis and you get to your patient very quickly. When you start mining the, the electronic medical record, you need to be, you need to be aware of the fact that in fact, you may actually find more patients than you have capacity to manage. And in the current environment where the, the therapeutic discoveries have preceded, you know, preceded our diagnostic and our regulatory pathways, you end up with a lot of patients who should be treated. You actually create a medical legal nightmare. They need to be treated. You've actually found them and you now have a ethical responsibility to respond. Did that answer the question? I, I, I hope so, but that's, it does to me, it does. The biggest challenge is you may actually find what you're looking for at a scale greater than you're anticipating. So can I ask you, all right, we got another question that is data mining makes sense and requires cooperation from all providers. Do you have support from payers or insurance industry? In the US, we appear to have inertia and discouragement of treatment with implement of prior authorization. Yeah, I thank you for the question. It's that, that's a universal, a universal problem and probably one of trust. And so it requires data sharing agreements and agreements very that requires the all parties to share data. And it comes down to whether the data is anonymized or identifiable. Identifiable data has more ethical restrictions, de-identified data for large, you know, if I've been able to access de-identified data earlier, I wouldn't have got into the mess because I would have actually known how many patients were actually identifiable and available. So that high level integration of data, which is de-identified, will give you a landscape, will allow you to do your site survey of what it is you're actually going to be mining. When it comes down to the extraction of the ore, except this time it's not the ore, it's actually the patient, then that is where the relationship becomes more challenging. But it's not, it's not insurmountable. There just has to be an incentive for the third party payer to actually want the thing to be treated. I agree. I think the health economics, I think the health economics will drive the, will drive the discussion, will drive actually the relationship. If the third party payer, the insurer, actually gets a better outcome by ensuring that patients are treated and cured from the disease, then there will be a better, a better outcome. And then what is the acceptance of AI in mining from your colleagues? So that will be in more, in a generic sense, AI in data mining, at this stage in the environments in which I'm working, it's still very, it's still very academic. And so colleagues are accepting of it because of the opportunities that it provides in terms of learning. The reason that the organisation I lead was established is because getting AI outcomes into practice has been, is challenging. It's always, it's always difficult to get it in as a sustainable practice. And so it is that there is a willingness to accept it and to, and to use it. But for osteoporosis, for example, it's already been implemented where the, where your radiology reports are now reporting vertebral crush fractures, incidental vertebral crush fractures. And it just goes to the primary care physician and they just hope for the best. Well, that's a Model E fractalisation service, which has not got any evidence that it's going to make a difference at all. So the technology is ahead of the regulation and the response. So it seems like there's no more question from our attendees. This was an excellent eye-opening talk about diagnosing more patients. And this is the same conversation that is happening in other field of medicines that for example, adrenal adenoma that may be reported in radiology but is not further worked up as. So the same thought is implemented there that if it's documented, there should be an advice further to it and how AI can help us identify those patients earlier. So as you said, be careful what you wish for, right? You might actually get it. I think it was Oscar Wilde or they said, there are two tragedies in life, not getting what you want and then getting it. And my journey was I got too clever for myself and I was almost, look, I probably shouldn't say this but I was almost relieved that the administration, you know, did a system upgrade and shut me down because I had created. And so unlike the adrenal adenoma experience where there is a question mark over, well, how, yeah, I understand. Here, osteoporosis, the treatments have accelerated to the point where you've now got a clear indication. Once you've identified the patient, there is a medical legal responsibility to respond because the treatment is known. And in fact, with some of the osteo-anabolics and treating to a target, it's actually curable. So in my career, we've moved on from not knowing how to treat it to actually knowing how to treat it. And AI now becomes both an asset and a liability. I agree with that. I have a question. The people who were diagnosed via AI who were not treated, was there any reasons behind it? Okay, I don't have data so I can only give you anecdote but when you, so when we started, we had better success at the start than as the waiting list got longer and longer and longer. By the time you're ringing people up 12 months after they've had a fracture, it's what fracture? Oh, she'll be right, it'll be okay. And so that's a very challenging conversation to have where you're talking to a person about osteoporosis that they had a year, a fracture event that they had a year earlier. And it just takes that long to get through that thousands of patients, right? And so to shorten the interval was going to require more investment from the hospitals where the fractures are being identified to the patient, but the challenge is then referring that patient back to primary care where the treatment needs to continue. There's a question there from Freddie talking about curing osteoporosis. So yeah, a patient with a vertebral crush fracture could now have, who has not been treated before, could now be treated with an osteoanabolic that largely takes their bone density from an osteoporotic to a near normal or to a higher level or actually normal for, and then holding those, that bone density at that higher level would to my mind be a cure. Perhaps it's just a treatment, but compared to the paradigm of just giving a therapy that stops bone from disappearing, the osteoanabolics follow the sequential therapy of osteoporosis is I think a game changer. I think that's the wrap up. Okay. Well, there's one more question. It appears that we need a town hall meeting to create collaboration. Yes, well, food always helps. I don't think you can get over the fact that we're so we're using technology to have a seminar that we otherwise would never have. Okay, so technology has its advantages. But the idea of a town hall is actually meeting each other and agreeing to move forward on a collegial and collaborative basis. Right. So there is that that town hall component of getting in the same room and getting on the same page. So, and not all communication is verbal. So I do agree with the town hall concept. And, yeah. And you work with the willing is my advice go with the people. I'm just going to go there. Yeah, go with the people who are going to implement a quickly and file quickly. So in some ways I was, as I indicated, we developed an AI tool for identification of fracture and couldn't actually manage what we'd actually uncovered. So that, you know, the next time we come back to it, I've got the data for exactly how many we're going to be uncovering. And you would need a facility to actually manage the, the, the, the uptake. I think the fact that the treatments of osteoporosis are now so robust that once primary care physicians see this as their, their primary core, you know, that it's as it's as logical to treat osteoporosis as it is hypertension and hyperlipidemia, which they do fantastically well. And they're comfortable with it. Then I think that will break down barriers and then the transfer of care into primary care will be more effective. So that's going to require town hall with a lot of people agreeing to proceed on that basis. There was a question there. I only get to see patients after the first, first of the fractures, but that's the time to intervene. That's when after the first fracture, their likelihood of another fracture is even higher. So it's secondary prevention is incredibly effective. There's another question too. Are patients told to talk to their providers about the findings that show need to treat osteoporosis is only the primary care provider expected to create or the end of who may be seeing the patients for another reason. Yeah. Thank you for the question. And to be effective, there has to be educated. You basically make the patient the advocate for their own health. So if there is no patient education component, then the, you know, the putting a fracture into a radiology report and sending it to the primary care physician and say here, manage this is asking for trouble. You're basically creating a medical legal minefield. It's you're putting a report. It's almost to my mind. It's almost like saying there's a shadow on the lung and, you know, do something about it. And if the primary care physician, you're wedging the primary care physician between having to do something and observing. So I find that I find that would be ineffective. I don't find that just notifying the GP is a satisfactory response. And all the work from curtain gander and others in that field insist on identification and education as some component of the of the of the fractionalize and pathway. Yes, Freddie, it would be Rama says mad for 12 months. And then once you've started it, it's done awesome. Well, it's some anti-resort given. Thank you for your time. Have I finished early? I didn't keep an eye on my timer. It's only a few minutes.

AI in healthcare

osteoporosis management

data mining

bone health

medical records

ethical standards

regulatory compliance

stakeholder cooperation

healthcare technology