iTranslated by AI
re:Invent 2025: Streamlining Clinical Documentation with AI and AWS HealthScribe: The Veradigm Case Study
Introduction
By transcribing various overseas lectures into Japanese articles, we make high-quality, hidden knowledge more accessible. In this project, we are taking on the following presentation!
For transcription articles of re:Invent 2025, please refer to this Spreadsheet where information is consolidated.
📖re:Invent 2025: AWS re:Invent 2025 - Reclaiming Clinical Time: How Veradigm Uses AI to Transform Workflows (IND208)
In this video, the AWS Healthcare AI team and Veradigm explain solutions designed to reduce the administrative burden on clinicians. Addressing the reality that clinicians spend two hours on documentation for every hour spent with patients, this session introduces implementation examples of real-time clinical documentation via AWS HealthScribe and secure FHIR data integration powered by AWS HealthLake. Through its introduction in Practice Fusion EHR, they have achieved time savings of two hours per day for each healthcare provider, while attaining a level of usability where 65% of users can start using it without training. It details technical innovations aiming for zero-edit note generation—including patient context integration, support for over 18 specialties, and AI safety guardrails—and how to build an AI-ready data foundation leveraging FHIR.
Note: This article is automatically generated while maintaining the content of the existing lecture to the maximum extent possible. Please be aware that there may be typos or incorrect information.
Main Content
The Challenge of Administrative Burden for Clinicians: 2 Hours of Paperwork for 1 Hour with Patients
Welcome, everyone. Good morning, and I wish you all a great Friday. Welcome to today's session. My name is Himanshu Joshi, and I am a product leader on the AWS Healthcare AI team. We are a team within AWS focused on building dedicated services for healthcare and life sciences customers. Today, I want to talk about something every clinician wishes they had more of. You already know what it is: more time. Not time spent facing a keyboard or using apps, but time spent facing patients and being with their families.
Today, I am joined by Mirza and Ken to discuss how AWS and Veradigm are truly transforming the point-of-care experience for clinicians through autonomous AI scribing and secure data foundations. By the end of this session, you will have a deep understanding of how the combination of a data foundation, dedicated AI agents, and a superior point-of-care user experience can genuinely transform the clinician's experience.
Now, let's talk about the problem we are trying to solve here. It is no news to anyone today: clinicians are literally drowning in paperwork. That's not the reason they went to medical school, right? Today, they spend a massive amount of time filling out forms, tracking clinical documentation requirements, and conversing with various stakeholders inside and outside their clinics to provide care. This leads to burnout and means they spend more time on tasks than truly facing their patients.
According to a very famous study by Forbes, for every hour a clinician spends with a patient, they spend two hours on clinical documentation and administrative tasks. This leads to a lot of burnout for clinicians, and they often have to spend time writing charts or responding to patient messages even after they return home. This contributes to burnout and is known in the healthcare industry as "pajama time." So, the big question for us is: what can we do to truly solve this pain for clinicians and allow them to spend more time with patients and their families?
Let's briefly talk about these administrative tasks at the point of care. If you look at the big picture of these tasks, they occur throughout the entire lifecycle. It starts before the visit. To grasp a true longitudinal picture of a patient, clinicians must review numerous PDFs while looking at an infinite scroll on the right side of the EHR. If a patient sees a specialist and returns to their primary care provider, they may need to reconcile a medication list.
At the point of care, they have to take notes in shorthand while talking to the patient, which sometimes does not lead to a very good experience for the patient. And even after the consultation is over, there are various tasks that follow, such as filling out forms for prior authorization and communicating with medical and billing teams. Today, generative AI offers a wonderful opportunity to transform this experience for healthcare providers. Let's talk about how.
AWS Approach: The Trinity of AI Data Foundation, Dedicated Agents, and Seamless UX
We at AWS strongly believe that solving the problem of administrative burden for clinicians is not just about the AI layer. Simply waving an AI magic wand won't make all these administrative challenges disappear. We believe in a combination of an AI data foundation. This allows every agent deployed at the point of care to truly grasp the patient's holistic context. And on top of this secure data foundation, there are agents specialized for healthcare. These are built truly for purpose to execute specific tasks well—it might be clinical documentation, past insights, or medical coding, but they are truly healthcare-specialized agents.
And on top of that, these agents and the data foundation must be provided at the point of care so that clinicians can actually use them. Because if the end-user experience for the clinician isn't seamless—if they have to switch between multiple apps to use the right agent—that's where adoption challenges emerge. But what are the challenges in advancing things across this entire stack? Getting patient context in a single format is not easy. It sometimes resides in the EHR, in multiple different systems, in PDFs, and is often not integrated in a discrete data format.
When it comes to generating notes at the point of care, there are challenges like multi-speaker scenarios, multilingualism, code-switching between languages—for example, between English and Spanish, commonly known as "Spanglish"—and multiple specialty scenarios. Many AI guardrails are required to ensure that the agents developed and deployed at the point of care are highly secure and earn the clinician's trust. Finally, it is essential to integrate these agents seamlessly into the clinician's workflow. By doing so, clinicians gain a truly seamless experience when deploying them.
We at AWS are innovating across the entire stack through partnerships with end customers like Veradigm and other EHRs to provide a truly differentiated experience. If you look at this diagram, there are three typical layers, which correspond to the three layers of innovation I mentioned earlier. AWS provides customers with a secure healthcare AI data foundation, through which you can save different formats of data into purpose-built data stores—such as AWS HealthLake, Health Imaging, and HealthOmics—depending on the data type. Clinical data can be securely stored in HealthLake, image data can use Health Imaging, and genomics can use HealthOmics. All these services operate at petabyte scale, are used by AWS tier-one customers, and comply with all rigorous security and regulatory-grade requirements regarding healthcare data and AI.
And on top of that, we provide customers with a suite of intelligent tools and agents. A prime example is AWS HealthScribe, which we will dive deeper into in today's session. HealthScribe is a clinical documentation service specifically built for generating clinical notes at the point of care. Customers can also use a combination of AWS Bedrock and Agent Core to create custom agents that simplify the clinician's life at the point of care. Finally, all these tools are available via APIs and SDKs, allowing EHRs and healthcare applications to integrate at the point of care and provide a truly seamless experience to the end-user clinician.
AWS HealthScribe Features: Real-time Clinical Documentation and Moving Toward Zero-Edit Note Generation
So today, we will dive deep into two of these services and discuss how Veradigm EHR is seamlessly integrating them into their workflows to provide an experience for clinicians. Let's start by diving deep into HealthScribe. Put simply, HealthScribe is a generative AI-based service that creates clinical notes from the conversation between a patient and a clinician. An easy way to understand what HealthScribe does is that when HealthScribe is integrated at the point of care or within an EHR, the clinician usually just needs to tap a mic to turn it on when the patient visit begins. HealthScribe's ambient capability listens to this conversation, and by the time the visit is over, it generates a highly accurate note based on the clinician's specialty and visit type.
I want to highlight three key features that HealthScribe provides to end customers. First is real-time clinical documentation. From the moment the clinician starts talking to the patient, HealthScribe generates a real-time transcript, and the clinician can keep checking it on the screen. Within seconds after the visit ends, an accurate note available at the point of care is generated, which the clinician can easily review and submit to the EHR. HealthScribe has built-in AI safety guardrails. For us, AI safety and guardrails are not optional. They are a mandatory feature for every AI agent developed by our healthcare team.
Digging deeper, every note generated by HealthScribe is based on source data. For example, look at the subjective section of a clinical note generated by HealthScribe. It includes references to the original transcript or conversation between the patient and the clinician, allowing the clinician to easily verify that hallucinations are limited and that the note is very complete. The final part is the point-of-care SDK. This is very important. We truly believe that the best experience for a clinician is one that does not require switching applications or changing workflows. Our APIs and SDKs allow HealthScribe to be easily integrated into any application, whether it's an EHR or an ISV-based application, so the clinician can simply click a button to turn on the mic and start talking with HealthScribe.
2025 Innovations: Building Toward an Experience Requiring No Edits
Now, let me talk about what we are doing with HealthScribe and what new innovations we are building in 2025. Our ultimate vision for HealthScribe is an experience where clinicians truly do not need to make any edits to the notes. From the moment you speak with HealthScribe, you can keep facing the patient and have a completely hands-free conversation, and by the end, a note is generated. Then, with limited or no editing, it can be easily reviewed and submitted. We have made numerous innovations to realize this. I will talk about some of them now.
New 2025 Features: Patient Context Integration, Dedicated AI Models, and Support for 18+ Specialties
The first one is patient context integration. In many cases, there are sections in notes that are already available as discrete data in the EHR—for example, patient demographics, pronouns, or even the problem list. With the new launch of patient context integration, when a new HealthScribe note is generated, customers simply add the patient context from their EHR or their own application, and we seamlessly weave it into the note itself. The benefit for the end-user is that they do not need to repeat the same information already residing in the EHR, such as demographics, specific allergies, or pronouns. This truly makes the path to a zero-edit experience a reality.
Next, what we are very excited about is that we also launched new dedicated AI models in 2025. We have fine-tuned these models specifically for clinical conversations at the point of care. What this enables is highly precise note generation control. You can truly control aspects like verbosity, completeness, and hallucinations within the note at the note level and at the aligned level.
The third very exciting feature is that we have launched five new note templates and tasks. We constantly receive requests from customers across physical health and behavioral health specialties asking if we can provide pre-generated templates within HealthScribe. And they want to avoid having to edit the note to fit the required format when the final note is displayed. This year, we launched five new integrated note templates, including SOAP, GIRP, BIRP, and SIRP. These are note templates for behavioral health specialties, and they are truly intended to enable an experience where these clinicians do not need to edit notes at all.
Finally, we are beginning to integrate tasks into HealthScribe. One of the tasks clinicians usually have to perform after a visit is writing a patient summary. This is often sent to the patient via email or printed material. In many cases, what we have seen is that the patient summary for the end-patient ends up being of a very medical or technical nature. Patients cannot easily understand it. So, what we are doing with HealthScribe is that when a customer sets a specific transaction or conversation to generate a patient summary, it generates a summary that the patient can easily understand. It clearly explains the patient's medications, follow-ups, and all steps, so the clinician does not need to write the patient summary, and it becomes very easy for the patient to understand the next steps of their care journey.
And finally, we have started support for 18+ new specialties. This is an effort toward our vision of achieving zero-edit notes. We have customers—large group practices—that support multiple specialties across their clinics. In such cases, they now love the experience of being able to support various specialties provided through a single ambient AI service. And finally, specialists need notes tailored to their specialty. For example, if you look at how a cardiologist and a dermatologist write notes, they are quite different. Now, all these specialty-specific nuances are already considered in HealthScribe.
Implementing HealthScribe at Veradigm Practice Fusion: Supporting 5 Million Patient Visits per Month with Ambient Scribe
Now, I would like to invite Ken to the stage to talk about how you are using HealthScribe in Practice Fusion EHR and the amazing clinician experience we are providing. Thank you, Ken. My name is Ken Sheppard, and I am from Veradigm. My background is as a software engineer. Today, I will talk about technology, products, and strategy—all of it. Let me tell you a little bit about Veradigm. We have three teams.
We have a provider team that provides solutions for healthcare providers and patients—that is, electronic health records, practice management systems, billing systems, and patient engagement systems. Everything a clinic needs to run its business. We also have a payer team that works with health insurers like Blue Cross Blue Shield, Medicare, and Medicaid to retrieve medical records for adjudicating claims and to create point-of-care alerts for clinics to participate in value-based care. And we have a life sciences business that leverages all our data to provide real-world data and real-world evidence, attempting to improve health outcomes through clinical trials. That is everything at Veradigm.
Today, I want to talk more specifically about our largest EHR, which is called Practice Fusion, and what Practice Fusion is doing with HealthScribe and HealthLake. Practice Fusion targets small, independent outpatient healthcare providers. It is not used in hospitals. While it is used in some large clinics, the sweet spot is really the 1-to-3 provider scale. It is designed to be modern and accessible. It runs 100% on AWS. Our mantra is "operational in 5 minutes." That means anyone can go to www.practicefusion.com right now, get a free trial, enter their credit card, start a subscription, and they get a real EHR. If you send us your medical credentials, we can enable e-prescribing and lab orders, and you can get started right away.
This is really important for reducing administrative burden, making it easy for people to use our product and get support. Again, because these clinics are small and independent, they don't have teams of personnel to conduct training, so it is designed to be very streamlined, very simple, and very accessible. And the Practice Fusion platform has a fairly large footprint. There are 5 million patient visits per month on the platform, and 660 million per year. As we speak, tens of thousands of people are having visits, getting e-prescriptions, and having their records documented on that platform.
Yes, moving on. As Himanshu said, for every hour a clinician spends on patient care, they spend another two hours on administrative tasks. The vast majority of that is chart documentation, but it also includes everything else, such as prior authorizations. And we have all seen that the computer in the exam room is actually a barrier to patient care. We have all sat at a table where the provider is typing away at a computer. That is clearly—and it has been like this for 20, 30 years—not the best way to care for a patient.
So, what we have heard from customers over the past two years—the latest being just a few months ago—was that the most requested feature in user surveys was ambient scribe technology. I have been working in EHRs for over 10 years, and I don't remember a time when providers were actually clamoring for new technology. Many of them still think paper is better, so this is a major shift. Furthermore, 60% of our users report using some form of ambient technology, which is phenomenal. Two years ago, the number wasn't anywhere near that. We weren't even asking that question two years ago. We asked slightly different questions about dictation and such, but the number wasn't near that. This means there is huge demand from users.
And from an ROI perspective, ambient scribe checks all the boxes. It saves providers' time. It helps create better charts. If you create better charts, you get better reimbursement. And ultimately, it leads to better patient care—such as being able to find clinical trials for patients, and so on. In other words, ambient scribe has a very high ROI.
Strategic Approach to Ambient Scribe
Moving on, we chose to partner with AWS to deploy an ambient scribe solution built on HealthScribe. I want to spend a few minutes talking about the strategy and reasoning regarding product fit. We believe ambient scribe is here to stay. We believe it needs to be an organic and essential part of the EHR. We also believe the charting experience will evolve with AI, and ambient scribe is like the starting point.
It is a foundation for connecting many other AI features. In some areas of our portfolio, we just bring in partners and they provide it to users. We still have, to be clear, scribe partners on the platform. We intend to keep scribe platforms on the platform. However, we believe we can provide an 80/20 solution. We believe we can provide a very simple, reasonably priced solution that meets expectations for chart quality and achieves zero-edit in the moment. That is how we evaluated the need for this hybrid build, buy, and partnership strategy.
Another challenge is product fit, but since this is Generative AI, there is an actual cost to run it. It is not like a SaaS application where you don't have to think much about how much compute you use or what the infrastructure costs will be. We determined that the only way to get market fit and build an economic model was to actually release something. That is, we needed to see how people were using it. HealthScribe is a single API call, so a very small team—smaller than a normal product team—was able to build and launch a secure, reliable scribe solution to market. By doing so, we were able to get this feedback, build an economic model, and understand how long consultations are, how many consultations are recorded, what the satisfaction level is with the note output, and all the other functionality.
This was very simple. Also, by building on the platform using an API instead of embedding it deeply from the start, we were able to move faster. This is working very well. We are releasing every two weeks on the Practice Fusion platform, but since the scribe is built on top of the platform, it releases even more frequently. AWS provides a tempo of continuous updates. We meet every week, provide a long feedback list from users, and receive those updates. We recently added support for pronouns, and some of the features on Himanshu's roadmap will also be reflected in the product soon.
The last factor that accelerated our time-to-market was what AWS has done with responsible AI. We are in the healthcare industry, so we have to be careful. If there is inaccuracy, there are actual patient safety concerns. There are concerns about bias and discrimination. Government regulations from the ONC require a lot of documentation to ensure safety and confirm there is no bias for predictive decision-support interventions, such as having a risk management program and model cards in place. Since AWS had those in place, it accelerated this process and got us 80% of the way there.
Furthermore, regarding AI, HealthScribe provides source attribution. You will see it in the demo; all the transcription is shown. We chose to display it side-by-side with the note and link them. This allows providers to trust the accuracy and cross the chasm of technology adoption. Because they can see, "This is what you said, and this is what was written in the note." We have received a lot of good feedback on this. Ultimately, this really is about us and why we chose HealthScribe. We were able to build an economic model, understand usage, understand how we could go to market, and get a ton of feature feedback, which we could continue to deploy.
Most importantly, what were the outcomes for providers? First, they saved a tremendous amount of time. It's almost unbelievable. They estimate saving 2 hours per provider per day. This is the result of combining surveys and analysis from providers using the system extensively. If you use it for only one consultation a day, no, you won't save 2 hours.
However, there is a lot of public data that scribes can achieve this. So, we really love that. The adoption was also very easy.
65% of users required no training. I don't mean human training, but watching videos or things like that. Most users literally just click to launch, hit the record button, understand what's happening, and when finished, press pause and hit save. That's it. This is great and aligns with our self-service philosophy. We received a lot of good qualitative feedback regarding focus on patients. With a 4.6 out of 5 rating, the scribe reduced charting time while with patients. We also have several quotes about that. This is exactly what we really want to achieve. We really want to enable providers to focus more on their patients. And there is a lot of other data that suggests patient satisfaction improves when using scribes, due to that personal care.
And we learned a lot about the note preferences of the providers who gave us a lot of feedback. There was a lot of satisfaction regarding the thoroughness of the notes. In some areas, providers have very diverse preferences. Some want bullet points, while others want narratives. Some want short narratives, while others want long narratives. Some want the patient's gender pronouns, while others want the pronouns stored in the EHR. Some want to refer to them as 'they', and some perhaps want to refer to them as 'the patient'. So, there are all these preferences that we have supported in AWS HealthScribe in collaboration with the AWS team, and we have already deployed many of them.
Now, let me share some quotes. This is what we want to see. Thanks to the scribe, I was able to complete 95% of my notes immediately after seeing patients. It is even better than other AI ambient scribes I have tried before. Again, the goal here—and I think everyone will reach it—is zero editing. Whether it is this month or next year, I believe all scribes, including AWS HealthScribe, will reach zero editing. Once that happens, we will be able to further expand the platform's AI capabilities. Here is another quote: 'The scribe has made me more efficient. It is what is called the HPI section and the plan or verbal discussion. I no longer have to type them. The scribe captures them well. I put them into my notes. Otherwise, when I am in the exam room, I have to type to keep up with them. Now, I just need to read back the notes from the scribe.' This is what we want to hear. And again, this is where we think we will reach 100%.
Demo and Future Outlook: Note Analysis Features and Patient Context Integration
Now, let me show you a demo. This demo is a bit rough, I will be honest. EHRs are very information-dense. This is also information-dense. I have expanded it to full screen. What you are about to see is a short video of launching the scribe in full screen. To make it easier to see, you will see where the visit is being transcribed—that is, the transcription and the evidence, wait, sorry, the transcription is the evidence—and the notes. And you will see the actually generated notes, which are then saved and reflected in the patient's actual medical record. So, I will click and move this forward.
Ah, it is playing. OK, it is playing. Great. So, you can see the transcription on the left. It has already started. I just pressed the save button. It takes a few seconds to generate the note, so on the left, you see the transcription with evidence, and this is the actually generated note, which you can save to Practice Fusion.
There are a few points to make. We built this simply because we wanted to iterate quickly. We have received a lot of good feedback on that. Also, this is the first time we have built something as a separate window within the EHR. Initially, we thought it might not be optimal, but in practice, we have received a lot of good feedback. This is because, for our providers, a big use case for the scribe is putting it on their smartphone and using it as a microphone in the exam room. So, having it as a separate window has worked well for now. As more AI features like coding and agents are added, it will become more deeply integrated into the EHR.
As you can see, there is not much display space, so we have actually received good feedback about the separate window. Now, let us move on.
So, what is next? We are working with AWS to develop an analysis feature that compares the notes generated by the scribe with the final notes. This is really cool for me and very exciting. There are not many features that can obtain this kind of quantitative data, so we will be able to measure things like how many times an edit was made, how much time was spent on those edits, whether it was changed because a drug name was wrong, or if a paragraph was converted into bullet points. And since we can run this on all notes, we will get near real-time feedback on whether we are meeting providers' expectations. It is also an excellent risk mitigation measure for safety, and if we start seeing many cases with zero edits, we can celebrate, right? That is what we are aiming for.
Therefore, we are working with AWS to build this pipeline, and we believe that ultimately, this data will be really good in terms of time savings and reduced editing. Well, let us see. As I mentioned, we will continue deeper integration into the EHR. We are rolling it out in its current form, and it is working well, and we are working on new features. And hopefully, in the near future, we will GA (General Availability) it for all users. Currently, we are limiting the users who can access it so that we can iterate on these features.
As an AWS HealthScribe partner, our intention is to release new features as soon as they are released from HealthScribe. That means they provide them to us as a beta, we test them, confirm they work for our use cases, and then release them. And what comes next, I think it relates to the quantity you mentioned, but what is quite exciting is patient context. Because, specifically, there are visits that do not work without patient context. For example, if there is a very short follow-up visit where the doctor asks 'How are you doing?', you say 'Oh, I am fine', and the doctor says 'Understood, if there is any problem, please give us a call', HealthScribe does not work. I mean, it does not work. There is no context.
However, by being on a path to inject context into that visit, we can understand that, 'Ah, this is a follow-up visit for shoulder surgery,' and generate appropriate notes. Therefore, I think ultimately, having context such as demographics and issues will allow for better judgment. I am very excited about that, and I am looking forward to it coming. Now, back to Mirza to talk about HealthLake.
Patient Context Challenges and AWS HealthLake: Realizing a Common Data Model with FHIR
Thanks, Ken. Can everyone hear me? Great. My name is Mirza Baig. I am part of the Healthcare AI team, the same team as Himanshu who was on stage earlier. Now, the point Ken mentioned last was about patient context. And patient context is clearly embedded in various different documents and different silos, which essentially hinders the true potential of all the AI agents being used.
Speaking of data silos, in this particular case, data exists across multiple tables. That means there are probably over 500 tables in an EHR or some relational database. Furthermore, data exists in various different formats. From the perspective of storage, publication, or even the format in which that data is specifically shared. For example, there might be CCDAs being shared, or data might be in a relational format and output as CSV. There are those aspects, but there is also the aspect of ontology, which adds an additional layer of complexity.
By ontology, I mean things like ICD-10 codes. For example, if HbA1c is referenced and there is a value of 7.1, what are the units? Is it milligrams, or what is the unit value? Was it done in a laboratory—that is, was it a laboratory with a specific CPT code—or was it a home device? Home devices have different CPT codes. Was it type 1 diabetes, or type 2? It has an ICD-10 association. Therefore, all these various things exist as additional layers on top of multiple formats and ontologies, complicating data reconciliation for analysis and making it difficult to make available from an agentic perspective. There is that layer as well. And as you all know, the vast majority of data actually exists in unstructured format.
This is incorporated into the notes, about which both Himanshu and Ken spoke regarding note generation. Ultimately, the next time a patient comes in, we need to bring the context generated in the notes into the next patient visit. Furthermore, it also includes all the various visit records that have occurred in the past. Hundreds of pages, thousands of pages of notes are essentially in an unutilized state, and there are various ontologies that are not only unutilized but are not clear enough to extract as chunks of information from an insight perspective. In other words, all these various complexities and various silos exist.
I would like you to focus on the center part of the screen, not the left side. From a technology stack perspective, what we decided was to focus on a common data model. That common data model is FHIR. This is not AWS's. It is open source. Basically, there is an aspect of a community that creates this specific common data model, and that common data model has an overall RESTful API architecture. What this means is, for example, if you have a FHIR API endpoint, I know that to get a patient's first and last name, there is a specific URL search string I must use to get that information. It might be something like a base URL slash patient, or if I am getting my own information, it might be patient question mark. This is consistent throughout. There is no need to reinvent the wheel.
In other words, it is a standardized data format that makes everything available, not just in how data is persisted, but also in how that data is exposed through APIs and through search parameters. One important thing I want you to understand is that this is all reference-based, not relational. What this means is, for example, suppose there is an observation result within a visit record. That observation result is related to the visit record. That visit record is, for example, related to a symptom, and that symptom is related to a patient. The patient has a relationship with a healthcare provider. The healthcare provider has a relationship with the organization to which he or she belongs. In other words, you can almost trace the graph in terms of creating these points of reference, these connections throughout. This makes it much easier to provide context.
Ultimately, it is not just about patient data or symptoms—though of course, it is about those—it is also about the healthcare system, the healthcare delivery system that realizes it, and all the various references available in this common data model are an inseparable part of it. Therefore, the healthcare-specific context and, of course, the schema parts are also quite normative. That is, for example, if I want to store LOINC codes related to checking my HbA1c, there is a code system and value set referenced when that assertion is made. That is, rather than a doctor simply saying or clearly entering into the EHR that they are performing this specific test, there is a reference in terms of what that value means when that test is performed. There is a LOINC code with a specific CPT assertion, which combined, performs that specific test.
And when the result is ingested again, the reference to that instruction via the code system and value set, as well as the unit value that emerged as a result of that specific result—all of these are stored in this specific schema, prepared and available so that models or agents do not need to think or guess what it means to extract that chunk of information. In other words, all of these are part of the overall standardized healthcare resources. On the left, there are, obviously, all these various use cases. You can plug HealthScribe into it. There are multiple agents; some are proprietary, some are first-party, and some are third-party. That is, there are all these various things.
This specific slide explains it from a fundamentally different perspective. Because we are not talking about a data model here. We are actually talking about how that entire environment is working symbiotically and seamlessly. On the left, I explained all the various different sources. There are more. I am sure you are familiar with them. In the center ribbon, there are various formats, and there are data transformation agents, partners, and transformation layers with necessary conversions. However, that data ultimately goes into HealthLake, and HealthLake also has integration with the MCP server, allowing it to perform queries, searches, and summaries of the patient's overall context.
Again, context is important. For example, the fact that Mirza has heart disease, such as cardiomyopathy, is significant from the perspective of the patient's visit, which the doctor and everyone else must be aware of when the visit takes place. Therefore, performing a summary and bringing that important information to the forefront as primary information, all of which is consumed at the top layer. These are use-case-specific AI agents. That is, I first talked about models. The second slide was essentially about how that model fits into this overall reference architecture. And this specific slide talks about the actual service itself.
The service I am talking about is HealthLake, and it has sub-second latency. This is very important, for example, in the case of Veradigm. Because ultimately, it is about patient care. Doctors and providers do not want to wait even a minute. Latency is very important for being responsive and from an adoption perspective as well. As you all know, it is built for performance and built for scale. It is petabytes of data.
For example, Ken mentioned that the Practice Fusion platform has 60 million visits per year. That is a very large number of visits. There is a lot of data created based on those visits, and whether you are doing it for 10 gigabytes of data or 100 terabytes of data, you need to be able to do it. That sub-second latency is very important. That is, performance; AWS is built for scale and performance. And of course, analytics is also part of it. That is, you can analyze the same data with Iceberg tables. Because as a managed service, HealthLake has this component, and that data is also converted into Iceberg tables so that it can be analyzed.
Veradigm's HealthLake Adoption Strategy: Building an AI-Ready Data Foundation with a Managed Solution at One-Fifth the Cost
We use Redshift and Athena, and we also incorporate third parties. We also use Databricks, Snowflake, and others to actually analyze that data and the resulting data, and make it available for AI agents as well. Now, I would like to call Ken back to the stage for the next few slides. Thank you. Yes, let me talk a little bit about HealthLake and comment a little bit on the problem of unstructured data.
As you know, healthcare providers are really busy and are under pressure in terms of reimbursement. As an engineer, you might imagine that you would want to put in as much structured data as possible for future use cases, but in reality, they don't tend to do that. They are busy, so they enter free-text notes. Obviously, AI unlocks that, and now NLP is converting notes into structured data, but this continues to be a major challenge.
So, let me start with a little bit of history on why EHR and why Practice Fusion is using FHIR and FHIR servers. I am sure you understand the ecosystem here. As you know, being a certified EHR means you can use that product to receive reimbursement from the largest payers, such as Medicare, Medicaid, ONC, and Tricare. They have requirements for the technology used, just as any business does, which is what is called a certified EHR. Every year or so, we must be recertified against new requirements that the government imposes on EHRs. Because they are trying to manage costs and improve health outcomes. It all makes sense.
One of the requirements for 2023 was to provide a complete set of FHIR APIs to app developers. That is, documented APIs, in FHIR, in JSON, with some OAuth flows—a patient flow, a provider flow—quite modern. In 2023, we deployed our first FHIR server using a partner to meet that requirement. I think 99% of Practice Fusion users had never heard of FHIR, and probably didn't know how to spell FHIR either.
Therefore, because our users wanted us to work on other things, we made a very small investment to meet the requirement. Now, two years later, these FHIR servers are being used extensively. Thousands of clinics have signed up. There are hundreds of apps, and 60% of the apps in our marketplace have some form of AI hook, so we are seeing more load, unpredictable load. Unfortunately, we see that current functionality limits what AI can do. Because we have to apply rate limiting and such.
For all these reasons, because the current solution is expensive, we considered HealthLake. HealthLake is about one-fifth the cost of our current solution. HealthLake is fully managed, so in a good way, it's really a black box. We have to build some software around it, but it works as expected. It is there, available, scales, and meets uptime, availability, and performance SLAs. Again, because it is scalable, our clinics can onboard themselves, and we can onboard clients, apps, and API clients. We can use HealthLake as an additional read destination, and ultimately help with the read load of internal workloads and decouple from legacy systems. Scalability is wonderful.
Another thing that is important to us is compliance. This is driven by the government. We must maintain compliance. Currently, I think we are at FHIR R4 US Core 6, but there are changes every year. Next year, there might be FHIR R5 or FHIR R6, but AWS is committed to supporting it. Therefore, we know we can maintain certification. We don't have to do schema migrations or deploy new hardware, so this is a great situation for us.
Again, for the sake of completeness, it is not just the government now. FHIR has become the absolute standard for interoperability. It is the standard for AI interoperability. It is already established. And looking forward, as Himanshu said, AI-ready data is difficult. Making EHR data AI-ready is very difficult. It's unstructured, deeply nested, and has all these relationships. You can imagine ETL-ing to other data stores, putting it into vector stores or graph RAG, doing all this heavy lifting, and keeping it up to date. It is difficult, but we believe HealthLake will make it much easier.
First of all, gathering data in one place, right? If you can gather all healthcare data in one place, that is a huge benefit. And due to some of what we saw in the roadmap earlier, and what we imagine will be realized by agents and MCPs, we believe AI will be able to utilize that data more easily. This is wonderful. Because otherwise, someone would have to extract FHIR data, put it into another data store, put it into a vector store, and do all the heavy lifting. Therefore, we expect HealthLake to enable AI innovation for us.
And going a little deeper, we have been talking about reading data, but HealthLake is a great place to store AI-generated data and AI inferences. In an EHR with a relational schema, no one was really thinking about AI inferences. Relational schemas are slow to work with and risky, but in HealthLake, if you are inferring a condition or an observation, it is an HTTP PUT or POST of that JSON element. You can mark it as AI-generated, and it becomes the source of record for that inference, which the EHR or other AI consumers and APIs can read. Therefore, long-term, it is our intention to start writing AI data there.
Is there anything else you want to say? This is why we chose this. Here is a very simplified architecture of what we are doing. It is a 'Hello World' type of thing, but on the left, we have the EHR. It has a relational database. We have DynamoDB, S3, Redis, and many others, but the main ones are relational, and we stream that data into HealthLake.
On the right, we have AI clients, interoperability clients, and our bespoke APIs. All of these can now read from AWS HealthLake without worrying about scale. This is a good state. This reduces the load on DynamoDB and other systems. Long-term, we want these AI clients to be able to write inference results back to AWS HealthLake. Whether it is a third-party solution, AWS HealthScribe generating more inferences, codes, etc., or our own Amazon Bedrock-based AI solutions, all of them can write to AWS HealthLake, and it can become the system of record for that data.
And the EHR can choose. It can say, 'I will render from AWS HealthLake.' Because the performance is good. It is called a lake, but in terms of performance, it is actually like a document database. The EHR can choose. Do we want to ingest this data? Do we want to embed it? What do we want to do? We love this vision of a future architecture as a way to make data AI-ready and provide more AI solutions to our customers.
Summary: Build Transformative Healthcare Experiences with AWS HealthScribe and HealthLake
Anything else? Now, what is next? We are in the process of migrating clients to AWS HealthLake. We have a legacy solution, and AWS HealthLake is running in production. We plan to migrate clients and manage bubble costs, etc. We plan to enable AI clients to use AWS HealthLake as the primary read store. That is, it will be enabled and recommended. We will do things like increasing rate limits to let them know that they can truly succeed on our platform and provide AI solutions to customers.
We plan to migrate additional workloads to AWS HealthLake. I mentioned that we have a payer business. The payer business distributes FHIR bundles and CCDAs to payers. Let's get them from here. Because this is actually very clean data. It also contains all security tags and restricted disclosure rules. Since these are all embedded in FHIR, we can distribute data in a way that is compliant. And we will lean into that again. We are not doing it now, but we plan to allow AI clients to use AWS HealthLake to save those inferences. Because I think that is very agile.
Yes, that is all the information I have. Now, back to Mirza. Thank you. Please stay on stage. We have two more slides. One is a recommendation for you to start building with AWS HealthLake. I would like you to use not only AWS HealthLake but also AWS HealthScribe. Regarding AWS HealthLake, briefly, basically, if you go to the AWS console and launch a data store, it will be ready to use from a test and development perspective within 30 minutes.
Similarly, you can go to the console for AWS HealthScribe. Basically, if you simulate a conversation, it will create notes for you, so you can start testing. There are also SDKs you can utilize. However, the point is that we are providing tools that not only allow you to capture accurate and relevant perfect notes but also to enable those notes in a wider context setting on a common data model. This is basically the combination of AWS HealthLake and AWS HealthScribe, which is what we are proposing here, and what Ken was talking about regarding Veradigm.
That is, a common data model, enterprise-grade FHIR resources, security, performance, and controls regarding responsible AI guardrails that I talked about with AWS HealthScribe. These are all managed services. This means you do not have to piece together various different components in the backend. You can literally just leverage the APIs and start using them immediately. That way, you and your team can focus on the higher-value proposals that you are essentially building. This is the important point I want you to take home.
And of course, you can also build transformative healthcare experiences by introducing a natural language query interface. And by the fact that all that data is available, you can also trigger additional workloads, such as calling other agents that you might be utilizing. So, there are two QR codes, so please capture them if you need to. If you cannot capture them, just go to AWS.com/healthlake or AWS.com/healthscribe, and you are ready to go. You can access the console or other information needed to start testing and using them from the product page.
Ken, Himanshu, any final words? I think we should answer questions before we wrap up. Yes, yes, we can be available for a few minutes after the session. As Ken said, we would be happy to talk to you. Thank you very much.
- This article was automatically generated using Amazon Bedrock while maintaining the information from the original video as much as possible.


























Discussion