Hello my name is Arun. I'm a Senior Technical Product Manager and I want to talk with you about Amazon Comprehend Medical. So what is Amazon Comprehend Medical? Comprehend Medical is a set of hyper-eligible ML powered APIs built specifically for the healthcare domain. It makes it easy to extract and structure information from unstructured medical text. It does this with state of the art accuracy, helping developers build applications that can improve patient outcomes. Comprehend Medical makes advanced medical text analytics accessible to all developers with no upfront costs. What problems exist that Comprehend Medical can solve? More than 90 percent of healthcare providers use electronic health records to store a patient data. Although there are few structured fields in an EHR, most of the valuable patient-care information is trapped in a free form clinical text. An example being admission notes, patient history, or discharge notes. However, extracting value define insights from unstructured clinical notes is a manual and labor-intensive process, creating a bottleneck that slows down analysis that could result in better health and business outcomes. You might be asking, how does Comprehend Medical actually work? Well, it is an extension of Amazon comprehends natural language processing models for entity extraction of medical texts. It uses deep learning to extract entities from unstructured text in the healthcare field such as clinical notes and radiology readings. Amazon comprehend leverages the latest advancements in machine learning to bring a high level of accuracy and efficiency to extracting clinical information. Comprehend Medical consists of two APIs. The NERA API, which will return a JSON with all the extracted entities, their traits, and the relationships between them. The second API is the PHId API which will return just the protected health information contained in the text. Developers can easily integrate Comprehend Medical into their data processing pipelines with tools like Amazon glue. They can also access it from SageMaker and extract structured data to build accurate models for health care use cases. Once the text is extracted, it can be stored in services like S3, Aurora, RDS, and Redshift or any third party service. What does this mean for you? Comprehend Medical can help improve outcomes. Identifying a high-risk patient on time will prevent further complications for the patient and reduce the financial costs for the health system. This data is also valuable for use cases such as clinical decision support, revenue cycle management and clinical trial management and it's difficult to use without significant manual effort. The ability to extract and structure information from unstructured medical text with state of the art accuracy no longer requires you to be a medical or machine-learning expert. Comprehend medical makes advanced medical text analytics accessible to all developers with no upfront costs. Comprehend Medical's current performance has been better than what we have seen in academic benchmarks. Well, how about we take a look at Comprehend Medical right in the console so you can get an even better idea about how it works? So let's take a look at Amazon Comprehend medical's console. Here we have a de-identified clinical node, I'm going to run it through the service. So here what we return is the actual insights from this analyzed text. You can see that the color is specific to a different entity type that we extract. Here the orange is basically for PHI, Protected Health Information. Here green is for medical condition and we actually have something called entity traits that, one is negation. So if for example, a patient denies taking some medication, that medication would be negated. Then we show whether a diagnosis is a sign or symptom. Why that's important is that when you work downstream, it's very important to have that differentiation in order to fit into a lot of workflows that exist with our health care customers. What we also do with relationship extraction is we tie the subtypes to the parents. So for example here, the test name is platelet count and the actual value is varied significantly. So you here you can see through the UI that it is connected to the parent. What we've done with this is, we want to make it really easy for customers to sift and sort through this data not through simple searches but more complex ones. So you can actually search, for example, medication also has relationship extraction. You can search dosages for medication. So what we do is, we tie the medication to the dosage route or mode, the strength frequency, and so you can make simple searches about how many of my patients are taking a certain dosage of a certain medication. In the end, we want to make this easy to use. We're really trying to distill a complex process into a very simple API call. Moving further, you can see another example of relationship extraction here under Vital Signs is temperature. So temperature is the test name and the value is 36 and it's connected to it. So it's very easy for our customers or for you to visually look at this and see how well our ML powered models are working. Moving further down in the console, you can see like we actually have a list type or we list all the entities and you can actually see the category and if any traits are associated. We also provide the conference score so you can see how well our models are performing on the unstructured data that you're running through the service. You can also sift and sort through this data as you deem fit. So essentially, if I wanted to look for example at test treatment and procedures, you can actually see the platelet count and under it, what the relationship was with the subtype and we provide that here as well. You can also close it. So again, we're trying to make this very easy to use and very easy for our customers to sift through. Honestly, very dense data very, very difficult to structure data. These will make it for you, the better you can use the service and obviously scale it for your specific use cases. Moving on to the JSON output, you can see here again, we make it as easy as possible and some of the identifiers that we have are ID, the offset like where it's located in the clinical texts, the actual score, the text and the category as I showed earlier, and the subtypes. So here the type for Labahn Hospital, it's under PHI, protected health information and the type is address. If you move further down you can see where we actually nest in the subtypes when we talk about relationship extraction. So now going back to the example of platelet count, you can see that that is a test name. It's under the category of test treatment procedure and the actual value below it is varied significantly. So again, when you take this data out of Comprehend Medical and put it into any database, it could be dynamo, it could be redshift. We want to make it very easy for you to search through it, and we believe that this format allows you to do that. With that, that's pretty much what the service does. We've really tried to take a very complex process and make it easy for our customers. We believe that if we can if we can do a lot of the heavy lifting and the hard work, we can actually enhance our customers to build all the really cool applications that can really change and impact healthcare. Thanks for learning about Amazon Comprehend Medical. On behalf of AWS, thanks for watching.
