Artificial Intelligence Offers Revenue Cycle Opportunities
Artificial intelligence will generate estimated savings of more than $150 billion worldwide for the healthcare industry, and abundant applications exist for the revenue cycle.
The use of artificial intelligence (AI) is growing across industries, health care included. In a new report, market research firm Frost & Sullivan estimates that by 2025, AI and cognitive computing will generate savings of more than $150 billion worldwide in clinical and operational areas of the healthcare industry. Meanwhile, in the revenue cycle, AI continues to bring benefits, such as improved efficiency and accuracy.
Eldon Richards, chief technology officer for Denver-based Recondo Technology, discusses how AI (including machine learning, natural language processing, and robotic process automation) is currently enhancing the revenue cycle and what is in store for future applications.
How is AI currently being used in the revenue cycle?
Richards: AI is being used in the revenue cycle in many ways. For example, hospitals are purchasing AI solutions, such as chat bots, that interact with patients using natural language processing capabilities. Such solutions also can be used for computer-assisted coding. Hospitals also are using AI-powered tools to generate claims. These are solutions that have been out for some time and are examples of technology investments in which purchasers know they are buying AI solutions. There are other tools where AI may be under the hood, making calculations or providing automation without the purchaser knowing the system is AI-enabled.
There are many opportunities within the revenue cycle to use AI to automate certain tasks, such as interactions between providers and payers. Some of these functions are done using older AI technology, such as rules-based systems. Increasingly, we’re seeing use of machine learning and natural language processing to enhance automation. For example, AI is being used to improve patient out-of-pocket estimates and predict denials before they happen.
How has AI improved the revenue cycle?
Richards: AI is improving accuracy for certain tasks and is raising automation rates for others. Machine learning is being used to optimize payer queries to ask for the right information and in as few transactions as possible while getting high-quality results. The new approach, using neural networks, versus the traditional approach of using rules to govern the process reduces costs by up to 25 percent without sacrificing accuracy.
Another area is out-of-pocket estimates. Traditionally, vendors have created estimates with a rules-based approach. They gather information from provider clients, create rules based on that data, and apply those rules to provide estimates. The shortcoming is that sometimes, individual factors impact what patients owe.
For example, a particular doctor may be using a different device for hip implants than the standard device a hospital is using. If that’s a consistent pattern, then machine learning, or neural networks, will be able to detect the difference for that doctor and accommodate for it, which we would never be able to do with a rules-based approach.
Authorization management uses natural language processing to automate the laborious process of understanding payer documentation requirements. It can be automated in part without AI, but using it produces the best results.
Computer-assisted coding has been in place for several years. It uses natural language processing to read documentation and generate the hospital claim codes.
Other AI capabilities include claims scrubbing, which allows providers to learn what edits are really required by payers, and scoring for payer appeals. Machine learning can identify which claims to focus on and which ones are likely to be overturned, so providers can focus their time on the claims that are most likely to be paid.
What are the barriers to using AI?
Richards: It is hard to find expertise in AI. There’s a big shortage given demand across industries. Another barrier is access to a sufficient amount of high-quality data to train AI models. In general, it’s an area that vendors are better positioned to do if they have permission from their clients to aggregate data from multiple clients to create a dataset that’s large enough to develop a good model. Frankly, only the larger hospitals and health systems are likely to have enough data to create good models.
What are examples of current AI applications that use high-quality data?
Richards: The truth is, all machine learning-based AI applications require high-quality data, meaning they need to learn from data that is accurate and relevant. For example, with patient estimation, the more obvious need is for accurate data. If a company creates a patient estimation model using inaccurate out-of-pocket actuals, then it will learn to make inaccurate predictions. The model is going to make predictions that closely match the data it learned on.
The less obvious need is for relevant data: The training data needs to come from a source that matches the production data with which the model will be used. Using patient estimation as an example again, if I create a model that only uses out-of-pocket actuals from a handful of my largest payers, that model most likely won’t be accurate with some of my regional payers, where my contracts might be quite different.
These two aspects of high-quality data hold for all kinds of AI models. I can’t create a high-quality lung cancer detection model if I only feed it images of brain cancer. Similarly, I can’t create a high-quality clinical natural language processing system by training it on Urban Dictionary data.
What type of money and time investment does AI require?
Richards: It runs the gamut. Investment in a purchased solution varies widely based on functionality. For the simpler cases, depending on hospital size, you could be looking at a few thousand dollars per month for a SaaS [software as a service] model.
Artificial Intelligence Improves Accounts Receivable Performance
Complex solutions, like computer-assisted coding, can run hundreds of thousands to even millions of dollars. Instead of purchasing solutions, hospitals may be looking at developing something internally. Those organizations are looking at expert teams that may cost $1 million or more per year. Implementation timeframes depend on the amount of data and the cleanliness of the data. If hospitals already have good analytics capabilities, they likely have large amounts of clean data. Those organizations may be looking at a three- to six-month project. If they do not, they can expect multi-year projects as they clean up the source systems that provide the data.
To summarize, the difficulty of a problem drives the complexity of the underlying model required to solve it. The complexity of the model drives the quantity of data and the level of expertise required to create and tune it. Predictive analytics and machine learning that can use a simple set of structured data can often be solved with simple models by an individual or small team.
Natural language understanding and computer vision problems require complex models and, usually, a large team of experts with a substantial amount of data. The cost to build these models depends on the team and data required.
What does the near future hold for AI?
Richards: The use cases I mentioned rely on AI functions that are currently available. Smaller hospitals could tap into these solutions if they are willing to create models by sharing data in a pool with other hospitals.
If you’re looking at what could be possible, assuming that the data could be pooled, then we could go to the next level. For example, with authorization, it would be possible to do full automation. The technology looks ahead at the documentation before an authorization is even requested to determine the likelihood that the supporting data will be sufficient for approval. If the technology determines that authorization will not be approved, it provides guidance back to the hospital on additional information that may be needed or other paths that should be pursued instead.
Similarly, there are tools that provide feedback on whether the data is clean, whether it is sufficient, and whether there is room for improvement. For example, clinical documentation improvement products could help identify gaps so hospitals can fill them and get more use out of AI.
Another application that’s probably more in the two- to five-year horizon is being able to learn from end-to-end revenue cycle activity sets. It changes how quality assurance is applied to the revenue cycle because we can look at what is generating denials or at situations where hospitals are not able to collect from patients. Hospitals can learn what data collected at the beginning of the revenue cycle may have been used to avoid denials or non-payment and improve the hospital’s ability to collect.
What should hospitals consider when implementing AI?
Richards: This is an exciting time. There have been many recent breakthroughs, particularly within the deep learning areas that open up many opportunities. However, healthcare leaders have to be careful with their AI decisions, particularly around vendor selection. It’s such a hot technology that there’s a lot of fake AI out there. Companies may have limited AI capabilities but are marketing them and can’t live up to their promises.
It’s important for buyers to do due diligence and look at what vendors’ capabilities are and how they are applying quality assurance to their models. Hospitals also should check vendor references carefully to verify healthcare expertise.
Karen Wagner is a freelance healthcare writer based in Forest Lake, Ill., and a member of HFMA’s First Illinois Chapter.
Interviewed for this article:
Eldon Richards is chief technology officer, Recondo Technology.