Colabora En La Factura De Desarrollo Web Para La Industria De Seguros Con Facilidad Usando airSlate SignNow

Web development invoice for Insurance Industry

- [Shashwat] Hello. Good afternoon everyone, and welcome to this session on automating insurance document processing with AI. My name is Shashwat Sapre, and I'm a senior product manager with the Amazon Textract team. I'm excited to be here today, and to present with Reddi Gudla who's the Staff Vice President of Digital Engineering at Elevance Health. - Thanks Shash. Go ahead. - So I'd like to start this presentation out by giving you a brief overview of Amazon Textract, and talking about some of its core capabilities, and introducing the concept of intelligent document processing. Next, I'll pass over to Reddi, who will then talk about Elevance Health, and how they're using AIML technologies, including Amazon Textract to solve challenges related to document processing in the health insurance industry. And lastly, I'd like to close it out by giving you some important resources and links that'll help you get started with your document processing journey. So despite the rapid digital transformation across organizations through cloud computing, and other technologies, documents remain a primary business tool in order to, documents remain a primary business tool across industries. Now, these documents can be in printed text, handwritten form, or both, but processing these documents is critical for our customer success. Now the insurance industry is no different. Here we can see some common types of documents that are present in high volumes in the insurance industry. These are documents such as patient enrollment forms, claims grievances, and appeals, invoices, and many, many more. Now traditionally, documents have been processed in one of three ways. Now each has its own unique set of challenges. The most common is Legacy OCR solutions. These are used by many different businesses and organizations across different industries. Now Legacy OCR solutions work well for simple documents, because they actually output a bunch of text, and strip the documents of their structure. Things like paragraphs, et cetera, into a bunch of simple text. And secondly, Legacy OCR solutions work best for documents that are in high or pristine quality. But when the quality of the input document varies, what we found is that the output performance of these Legacy OCR solutions can vary, and so typically doesn't meet our customer's needs. The second method is actually manual processing via human review. And as you can imagine this process is time consuming, it's expensive, and because there are humans involved, it's also error prone, right? Humans can make mistakes, and so there's a degree of variability in the output. And these mistakes then need to be accounted for, and are an extra cost that needs to be accounted for. Now the last method is rules and template based extraction. Now this method is only as good as the OCR technology it's built on. But this method also tends to be brittle, and difficult to scale because there's an overhead involved in managing all those rules and templates for all the different types of documents that you're going to wanna process. If any of those documents found fall outside of the rules and templates that you have set up, it actually causes errors and again, needs to be accounted for. But regardless of whatever method you choose, the goal of document processing remains the same, and that is to get the critical information from your documents into your decision making systems or people as quickly and as accurately as possible. So let's take the example of insurance claims processing right. Here, what we're trying to do is get the dozen or so data points from our claims documents as quickly, and accurately as possible so that we can actually drive a decision on whether or not to approve or deny a claim. So to address the challenges of legacy document processing, we actually built Amazon Textract, which enables intelligent document processing. So what is intelligent document processing, you might ask. Well, using intelligent document processing, we can actually shorten decision cycles by giving you higher quality data, and that is more accurate. Enabling you to make faster decisions. Now this in turn allows you to serve more customers, because you're able to repurpose your existing workforce to higher value tasks that only people can do. Now let me switch gears and tell you a little bit about Amazon Textract, and how it can transform your business processes. So Amazon Textract is a managed machine learning service that extracts printed text, handwriting, and data from virtually any document. So it essentially uses AI to transform documents into useful structured data. And this all starts with accuracy, because customers have told us, and actually you've told us that the most important thing when processing documents is accuracy. But we don't just think of accuracy in terms of outputting a bunch of text on a page. We also need to think about the relationship of data in a document, and the data, the structure of the data within the document. So Textract actually does this. It goes beyond traditional OCR to extract the text, handwritten data, and data from documents while preserving data structure and data relationships. Next, Textract is actually, you know, built with security and compliance in mind. So from a security perspective, we have encryption at rest and during transit. So you can be assured that your data remains secure at all times. And from a compliance perspective, Textract is actually HIPAA compliant. So you can even use Textract to process documents containing PHI or protected health information. The great thing here about Amazon Textract is that you don't need to be a machine learning expert to use it. All you really need to do is integrate with our existing ready to use globally available APIs to get started. Furthermore, Textract is pay as you go. So there's no being locked into complex contracts, or getting involved in licensing fees, or complex negotiations. You're only billed for the pages that you process. And being on the AWS Cloud, we can leverage AWS's elasticity to quickly scale up and down depending on your demand or to meet your needs. So all of these benefits basically help to reduce time to value, right? Textract is completely serverless, and so there's no models to build or maintain. There are no templates, and there's no infrastructure to manage. I like to joke that we wanna make it as easy as possible for you to use Amazon Textract, and all you really need is a credit card and a developer. So let me tell you about why customers are actually choosing to use Textract and some of the things that you would need to consider if you choose to go the DIY route. That means to build your own ML models. And it all starts with data, right? Because data is a lifeblood of any good machine learning model, but this data doesn't just appear out of nowhere. You actually have to go and source this data, then you have to label or annotate the data, and then you have to transform the data. So it's in a usable state by machine learning models. And then now we get to models. So here you will actually need some machine learning expertise. You probably need a scientist or two who can help you to understand which machine learning model to pick for the data that you have. And then you need those scientists to actually train those models on the data that you've collected. And by the way, the training of the models itself is an involved process, it's a time consuming process, and once you get it right and you achieve those satisfactory or good outcomes, you're still not done. You have to then think about things like model deployment, model versioning, and then the ongoing model maintenance that you're doing. Now compare and contrast this to Textract, which as I mentioned is a managed machine learning service. We take care of all of this for you. Again, with Textract, there's no models or templates to build or maintain. There's no infrastructure to manage. All you need to do is get a developer, a credit card, integrate with our APIs. So this slide here basically shows a handful of public customers that are referenceable, but overall we have thousands of customers across different geographies and industries who are using Textract for a wide variety of use cases. Next, I'd like to switch gears, and just talk about some of the core features, or capabilities of Textract. Now, naturally we support OCR, right, or basic text extraction. But in addition to that, we also support handwriting extraction, and extraction of higher or more complex data structures like entire tables from documents, as well as key value pairs from different types of forms such as financial statements, loan applications, insurance claims, and more. Secondly, we also have support for specialized document types like invoices and receipts which come in all shapes and sizes, or even identity documents such as passports and identity cards. Then last year we actually launched Textract queries, which allows you to specify and extract only the information you care about from virtually any type of document without having to worry about document structure, which is huge. And lastly, our newest edition is signature detection, which as the name suggests, allows you to detect whether signature is present in a document, whether that's an e-signature, a handwritten signature, or simply it's initials. Now let's take a deeper look at forms and tables extraction, which is where Legacy OCR solutions have historically not performed very well. Now on the slide up here, what we have is an example of a CMS 1500 form, which is a standardized claims form that is used by health insurance providers to build Medicaid carriers. Now, Amazon Textract is able to extract all the claims information from this form with high accuracy while preserving the data structure and relationships. So if you look at the first box here, you can see that Amazon Textract is able to extract all the fields that are present in the form as key value pairs preserving the relationship between the field name, and the actual answer within it. So over here we can see the insured ID number, and then the series of ones actually represent the value of that key. Similarly, Textract can do the same thing with check boxes. So you can see that check boxes in this field, we can determine whether the married field, or determine that the married field is selected. Some of the other check boxes are not selected. Now handwriting is another thing that we support. So in this example, the patient actually signed their name, name in the signature field, and you can see that that itself was extracted accurately, and again displayed as a key value pair to maintain that relationship. And lastly, towards the bottom of the table, sorry, the bottom of the form, you can see that there's essentially a tabular structure or a table. Now Textract is able to extract the entire table, maintaining the rows and columns, and also providing the actual cells accurately so that you can perform easier post-processing later on once you have your output. Next I'd like to talk about an example using specialized document extraction. Now, invoices and their receipts are traditionally different document types to process at scale because they don't follow any set design rules, and actually you guys see a bunch of these different types of invoice and receipts in your normal course of business. But using Textract you can process virtually any type of invoice and receipt at scale without any templates or configuration required. Now in this example of a medical insurance invoice, what we can see is Textract is able to output the information accurately, again in key value pairs, but it outputs them in two kind of groups. First, you've got your summary fields such as the invoice number, subtotal, discount, tax, and total, as well as your individual line item fields like the full body checkup, the infection due to inflammation, et cetera, et cetera. Now using this, this essentially makes post-processing for you very easy, and it enables at scale processing of invoice and receipts of all shapes and sizes. And Textract actually goes one step further. It even normalizes those field names so that you can consolidate data across different formats or types of invoices and receipts. Now lastly, before I hand over to Reddi, I'd like to tell you about Textract queries, and why it's such a powerful tool to extract information from semi-structured and unstructured documents. So as a reminder, Textract queries allows you to use natural language questions to extract specific pieces of information from documents without worrying about the document structure at all. Now, we've heard from customers that basically they want to easily extract information from documents, but typically you don't care about all the information in the documents, just certain pieces of information in those documents. Paystubs are a great example of a semi-structured document, because as you know, they come in all shapes and sizes, but they always have a few common things. The name of the person whose pay stub it is, how much they're getting paid, so on and so forth. So here's where Textract queries really shines, because it allows you to just create these natural language questions, and essentially ask Textract what information you're interested in retrieving. And then Textract queries actually scans the document, and returns the answers to your questions in the form of an output. So in this example, I've got an insurance card, and what I'm interested in finding out are some things about, you know, who's the member name, what's the member id, insurance provider, and so on so forth. And we can see that Textract queries is able to accurately find the answer, and return it in the Textract of queries output. But what's interesting if you look carefully, is that you don't necessarily need to have labeled the thing that you're trying to find explicitly within the insurance card or the document for that matter. So if I look at member name, or insurance provider, or the plan type, you and I can tell where they are very easily by looking at the insurance card, but a machine learning model doesn't really know that it's there, right? These items aren't labeled at all, but Textract is still able to identify them, and return the right answer almost as if by magic. The other thing that's interesting to note is that Textract also supports common abbreviations. So I can write what is the SSN or social security number, and it'll know what I'm talking about. And that's exactly what happened in this case. Textract understands that when I write out of pocket maximum, it refers to the OOP over here, and it actually returns the right answer. So with Textract queries, you can essentially unlock the automation of processing of a higher proportion of workloads than was possible before, and actually automate the processing of documents that were traditionally very challenging to process using, you know, document processing technologies. Now with that, I'd like to pass over to Reddi to talk about Elevance Health, and how they're using Amazon Textract. - Thank you Shash. Thank you Shash. I'm Reddi Gudla. I lead the exponential engineering organization at Elevance Health. A quick overview of who we are. Elevance Health, we are formally called Anthem. Essentially we provide health insurance coverage for about 47 million members through our different affiliated health plans. Essentially what that means is coverage for one in eight members, is policies for them is coming from Elevance Health. Our mission is to simplify healthcare, and our vision is to be the most innovative, and the valuable player in the space. So let's look at our current business and the challenges. As you all know, the health insurance costs and healthcare costs are rising. And along with that, even the health insurance costs are increasing as well. And Elevance is making every attempt, every attempt to manage the cost. So there are different levels we can follow. One of the things is, how do you manage the operating costs, right? By increasing the efficiency of our operations, we are now looking at somewhere to manage the operating cost, right? And leveraging newer technologies, the AI and the machine learning tools, digitizing our operating for processes, the business workflows end-to-end, simplifying our data quality, increasing the data quality, providing bidirectional data. These are some of the technical opportunities we have that can help manage our operating costs, and digitize our end-to-end workflows. However, just like most other large enterprises, we do have the challenges with thousands of disparate data sources, multiple data formats, and more importantly, the need to have extensive domain knowledge. You need to be in the business to understand our business, to even understand the data model. So this is a people issue, and that's what subject matter expert issue. So all of these are limitations that are constraining our ability to digitize, but again, this is an industry problem. So our journey, right? I'll quickly talk about our journey in the AI space. We started the concept of digitization, or trying the process of digitization for over four years. However the pandemic lockdown, particularly the lockdown timeframe, I think this is around March, I think right after the lockdown happened. We have a situation where we have to expedite the digitization, the end. We have to increase their option of digital tools just like we do have, right? Our operations, we still have, we still receive a substantial amount of paper documents. And besides that, we also have electronic communication, both Word documents, Excel documents, but all of them require human intervention, which essentially means our cost of operations are going high. And between the various applications supporting all of the different stakeholders, the stakeholders like the members, the providers, the employees, all the applications supporting them, we have multiple use cases which are impeded by these limitations in getting more paper as well as the human intervention in our operations. So this resulted in the need for more digitization, and we also defined the need. There is a need to create a technology platform to enable this digitization. For this digitization, as we have, through our journey over the years, there are two core capabilities needed. One is the intelligent document processing, our ability to meaningfully extract, basically parse the content from different documents, images, scanned images, and extract meaningful content out of it. By meaningful content, content that's relevant within the context of health insurance domain. The second use case of importance to our second feature that we really need is a simplified semantic search, as applicable to the health insurance. How do we do that, right? The very first step in our journey before we looked into any further tools is to build a terms of reference, which can call it as dictionaries or ontologies, taxonomies, whatever's the right term used. But essentially what we have done is we spent enormous amount of time within our subject matter experts to pull together terms of reference of importance for each of our different domains, whether we call it the claims, right? The benefits, co-pays, coinsurance, deductibles, these are the different terms which are very unique to health insurance. So we kind of built a standard terms of reference. Once we had this terms of reference, we are now combined, we now looked at technology tools, right? Different technology tools like the OCRs, the NLPs, the AIP, and the machine learning technologies. Combining this with the terms of reference, now we built a core foundational platform, or cognitive ops platform, which is something what we can use for digitization of all the use cases. Basically to streamline the end to end workflows. So what you see right now is the two forms of our platform architecture to help digitize all the use cases. So the top part of the diagram, right there, which is reflecting to our most common use case. This use case relies on the data extraction pipeline. There's over I think seven or eight use cases tied to this kind of a pattern. So let's take a look at one example of how this pattern is used. Say a member goes to an out of network provider, out of network doctor, gets a service, pays for the service, collects the receipt, then wants to submit a claim to Elevance to get a reimbursement. So the member can, through our different online channels, as well as through our fax or mail channels, there are different channels, can submit this request along with the receipt. So once we receive the receipt, it goes through our data extraction pipeline, and this is where we use AWS Textract. There are previous versions of our solutions used, different versions of our other capabilities, but this is one of the ones we have used AWS Textract extensively now. What using AWS Textract, we do extract name value pairs of all the relevant contact with content within the receipt. And this information, the name value pairs along with other contextual information that we received, is now sent to our downstream processing for subsequent enrichment. And that is, that's a critical part. Just the OCR by itself or the extraction by itself doesn't help solve the end to end business case, right? The workflow for us to solve the end-to-end workflow, we do have to enrich the information. So using some proprietary machine learning models we built in other third party tools, and in some cases some plain old Java tools, we enrich the data that's extracted. For example, if we take an example of the receipt in this particular use case, there could be procedure codes, diagnosis codes, scattered all over the receipt, right? There are different kinds of receipts. You'll probably get the receipts with information in tables, in different columns. You should be having numbers all over the place. But some numbers can be relevant to our, could be a procedure code, some can be a diagnosis code, and it could be written in different forms. The ability of us to enrich, extract from the extracted content from Textract, to be able to translate it using some of our own enrichment, and identify what is the relevant procedure code, diagnosis code. I think that's where the enrichment comes in. So using this enrichment, we now have, probably have all the necessary information to now submit the claim to our claims process system to submit the reimbursement. So this is one example of this use case, there are multiple patterns. For example, there is one pattern with respect to grievance and appeals routing, right? If a grievance or an appeals case is received, how do you route it to the right agent? Okay, by extracting the right piece of information out of the extracted documents do an enrichment, we now are able to identify, okay, whether it should go to our government business, whether it should go to our commercial business. So many use cases like this, following this data pattern, using this pattern of extraction of data, and then subsequent enrichment helped solve some of the business end-to-end business workflows. Basically digitize the business workflows. The second pattern, which we have here, this is the use case, let's talk about this. This is the use case using our semantic search, right? The simplified semantic search engine. In this case I'm gonna talk about a use case we actually worked with AWS. This is what we're calling as a reimbursement methodology use case. Let's talk about an example here again. A provider, say a doctor, receives a payment from Anthem, say $80. But the provider is expecting say $120 for the service. So they do submit an inquiry to Elevance asking like why? Why did you, what's the difference? Can you provide an explanation? So when the inquiry received by our analyst, we do have to pour through multiple documents. Typically these are state sponsored documents. So we regard the state documents. There are multiple documents, hundreds of pages, and we have to sift through these documents, and any analyst has to look through, find the corresponding fee schedule associated with that. And if the fee scale, based on the fee schedule, they'll be able to answer the query. Typically takes anywhere from seven to 15 days to respond to an inquiry like this. So how do we plan to solve this? The idea there is to use Textract, and using the terms of reference with all the health insurance terms we have identified as the ontologies and the dictionaries we used. Using the terms of difference, reference, and using Textract we now parse through the documents, identify all the relevant content, and now we organize them into a meaningful knowledge model for faster retrieval through a search bar. So this is a POC we're working, I think the initial states of the POC with AWS are very promising. We're looking at, there are some more things we have to do, but very promising of that. But this pattern of a simplified search pattern, we have actually implemented with some of our other use cases especially when it comes to answering a inquiry from a member. If a member inquires, and ask us like if a wheelchair is covered. In a typical situation, you do have to, I think most of your enrollment, you might have got a hundred page document. It's called an EOC, evidence of courage document. If you have to find, like if you benefit is covered, you do have to read the hundred page document. And we did solve this using the semantic simple search button but not using Textract, but using some of our cognitive ops platform, and ability to extract the content, and organize it in a knowledge model for simplified search. So bottom line, this is a pattern which is promising, and using Textract, we are really looking at making it a bigger model for more of those use cases. So moving on, how did we pick, right? How did we pick the right AI solution for document processing? As I mentioned, right? We started this journey, started this journey of digitization to streamlining operations for all four years, and we started off with primitive tools, When we started off this journey, 346, there's mostly open source tools, there's some tools from Stanford core NLP, something like that. That's pretty much what we got when we started this journey of digitization. And over a period of time we have identified certain key requirements that are required for digitization. Obviously the first thing we know of is to build the terms of reference, which we have done. And the next thing is like use some of the good technologies out there. And through that process of the learnings over the last couple of years, we identified about five or six key capabilities that are needed. The very first one is obviously the very simplest one, Textual OCR capabilities with the confidence indicator. Second thing is coordinate mapping. You will probably need to know the coordinates, especially when you are extracting content, and are trying to make sense sort of the content from different forms, so coordinate mapping. The third and four are probably the most unique that came through our experience. This is about reading tables, reading tables, extracting content out of tables. More importantly, there'll be different kinds of tables around the different kinds of documents. The key thing you'll realize though is there are borderless tables. There are tables that span across different pages where the heading is in page one, but the rest of the content is in page two. How do you make sense out of it, right? Those definitely, those are definitely our edge cases. But the challenge is these edge cases do complicate the solution, and thereby increase the cost, or the accuracy of the solution. In some cases, it'll even prevent that option of the solution, the trust of the business to use the solution. So it's important that these are some things you learn through experience. These edge cases have to be solved. So page alignment, page numbers, form detection with key-value pairs, especially every year at enrollment, you'll start seeing so many new forms, and each time there's a new form, you do not want to be in a position to redo, or do more code changes to support that. So that's another key requirement. And lastly, the simple data type detection and conversions. So from these six core capabilities are the key requirements. About two years ago, just about when we had to expedite our usage of digitization, we did evaluate multiple products, and AWS Textract ranked higher than the rest in terms of all these core features. And more importantly, being in healthcare space, we do have to follow the tight security and compliance guidelines. And Textract did follow, it did follow throughout, did get our CSO's approvals. So we are glad to use that. And then once we decided to use Textract, the one thing that really make this is a good partnership with AWS, they're easy to work with at the time. I mean even now I would say, and open to feedback they, provided faster turnaround. Bottom line, we were able to solve many use cases, digitize a lot of our operations. I mean about six or seven use cases over a year, a year and a half timeframe. Okay and a few value adds, like how does Textract, and the overall end to end workflow automation and digitization help us? The top two use cases are examples of the first pattern, the data extraction pipeline. The first one you talk about is like how do you classify the correspondence documents for providers for claims processing? By doing this where we have quantified an operating gain about savings, about $500,000 annually. The second one is about intelligent routing of a grievance and appeals request to the right agent. Again, this one we're projecting a $9 million savings within the next five years. So again, there is many use cases. These are the two which we had some quantifying operating value identified. In the last use case. This is the one, I would say this is an advanced POC stage. This is the one I was referring to with the reimbursement methodology where a provider submits a request, and we are supposed to respond back. In current stage, it takes anywhere from seven to 15 days. And our goal is to go through a simplified search bar where you should wear the provider itself. We should be able to give to the provider itself, and the provider should be able to identify, and answer the questions themselves. So what next, as I mentioned, right? I think the key thing to digitize. Obviously is to build the terms of reference, and identifying the right amount of content, right? The right terms of reference as applicable to the health insurance. And then into the end of the medical terminology. We do want to expand our domain specific models using AWS Comprehend. The second one is to co-create solutions with AWS using a custom form recognizer. I think this is the situation where every year at enrollment period, we do have changes in the forms, in enrollment forms. In the current way, we do have to, for the enrichment process that we are following through, anytime there's a change, we do have to change some of our coding pattern, our coding solutions to start to do the enrichment. Working with AWS, we are hoping to change the custom form recognizer, thereby we can limit the change. One example we can say is, as we are doing the end-to-end digitization, I would say there is about 90% accuracy at this point from Textract subsequent enrichment to be able to submit a claim. In the case of the manual claims example, it's about 90% accuracy, and there's about 10% human in the loop. And then we are hoping as the AWS Textract, and some of these tools and with our own homegrown solutions and enrichment expanding on that, we are hoping to increase it beyond 90%. And lastly, to enrich our cognitive apps platform with the AWS A2I. This is the human in the loop interface. We currently have a homegrown interface for human in the loop. And we are looking at utilizing, probably leveraging Amazon A2I for that. With this, I'm going to turn it over to Shash and thank you. (crowd claps) - Thank you Reddi. (crowd claps) Thank you Reddi. So hopefully now you can see how powerful Amazon Textract can be to enhance your document processing workflows. Now with this, I'd like to move on to kind of talk about how you can get started with using Amazon Textract. So we have a partner ecosystem that can help you with the acceleration of your document processing journey. Now these aren't all the partners available, but this gives you a broad idea of how many partners there are. And the fact is that all these partners are actually experienced with our AWS platform and with our services. So they can help you in your digital transformation journey no matter what stage you're in. And lastly, I'd like to leave you with some important resources and links that can help you get started. So we've got a workshop that's online that you can go to that'll walk you through how to use Textract. We've got blog posts that talk about, you know, start to finish implementing Textract, as well as source code links if you want to jump straight in. And then as always, you can reach out to us to set up immersion day trainings, or to get any AWS partner referrals. And finally, I'd like to thank Reddi for presenting with me, and helping me tell the story about intelligent process, intelligent document processing in the insurance industry. Thank you everyone. (crowd claps)