RPA, AI help speed review of Medicare claims
Connecting state and local government leaders
The Centers for Medicare and Medicaid Services is using the Intake Process Automation Tool to validate information on Medicare Advantage claims and identify potential problems for staff reviewers.
Employees and contractors at the Centers of Medicare and Medicaid Services spend countless hours every year reviewing thousands of medical records to ensure the accuracy of Medicare Advantage payments. An automated intake tool is working to change that.
Using emerging technologies such as robotic process automation, optical character recognition, machine learning and artificial intelligence, KPMG’s Intake Process Automation Tool ingests records as they are submitted and identifies potential problems according to set parameters, submission rules and coding guidance. Specifically, RPA orchestrates the steps of the intake process, OCR digitizes the scanned document and then AI and machine learning are applied to understand the document and extract the information necessary to validate the information.
Intake PA stands to save CMS time and money, said Payam Mousavi, KPMG’s lead director for intelligent automation for governments and the technical lead for the CMS project. The tool processes a record in about a minute, whereas human reviewers average about 65 minutes per record. The agency received 40,000 records in 2018 and expects to get 50,000 this year.
To submit records, health care providers scan them in, resulting in varying formats and lengths. “It could go anywhere a few pages to over 1,000 pages,” Mousavi said. “Because they’re scanned in, there’s sometimes handwriting in them.”
In the past, the CMS staff reviewed all that unstructured data, making sure the scans of records were rotated properly and going page by page to check that they had the right date of service, provider and beneficiary.
Now, OCR handles the digitization, but to make sure that names, birthdates and other identifiers within one record match, KPMG created microservices that use natural language processing and machine learning to distinguish between a date of birth and a date of service. Once that data is in a structured format, it’s possible to do the validations automatically for each page, Mousavi said.
“If you look at a medical record, you may have multiple names in there,” he said. “You may have the patient name, you may have the physician name, you may have a nurse name, you may have a street name. To a human, it’s pretty easy to differentiate between them, but if you have a machine looking at it, that’s where we’re using natural language processing and machine learning to understand the document like a person would.”
Much of the tool is new technology that CMS did not have, Mousavi said. His team saw the highly manual, repetitive process as a good use case for RPA, but RPA is not ideal for unstructured data or making decisions, he added. That’s where AI and the KPMG Ignite AI platform come in. It has natural language processing and machine learning algorithms to help understand the unstructured data in the documents.
Besides time savings, Mousavi sees two other benefits to the tool. The first is cost savings because CMS doesn’t have to pay employees or contractors to handle the repetitive tasks and PA Intake can run 24/7. Still, the tool doesn’t completely remove humans from the review process. If the tool doesn’t have high confidence in the decision it made, the record will go to a person for verification.
“There are still some steps that cannot be automated or haven’t been automated yet or regulations don’t allow them to be automated,” Mousavi said. “But the expectation is that they’re going to get done faster.” And that translates to faster feedback to providers and better job satisfaction because workers are looking only at complicated records, not checking all the information on every single one.
Another perk is increased accuracy. In fact, the tool returns 95 percent accuracy, and a test found cases in which the bot discovered something invalid that a human had overlooked. “If you’re looking at 1,000 pages, it’s easy as a human to miss some of these things,” Mousavi said.
Since the PA Intake went live in April, it has processed more than 6,000 records. Currently, CMS is using one tool, but it hopes to have multiple bots up and running to speed the process further. In the fall, KPMG expects to add more validations such as whether the forms have signatures.
In May, the tool earned KPMG a spot as one of eight finalists at ACT-IAC’s Igniting Innovation 2019 conference, an event highlights the newest technologies in the public sector.
Editor's note: This article was changed July 2 to correct the spelling of Payam Mousavi's name.
NEXT STORY: Protect the border with tech, Hurd says