LEARNING OBJECTIVES
- Distinguish between the inflated expectations and the current reality of AI in health care billing.
- Understand the different layers of AI technology and their associated cost implications for implementation in RCM.
- Identify both successful applications and current challenges of AI in health care RCM.
Artificial intelligence (AI) has been presented as a force capable of transforming health care billing by optimizing processes, reducing costly errors and boosting reimbursements, especially as traditional offshoring strategies reach their limits. Some have even predicted that fully automated billing systems will eliminate the need for human personnel in a few months. However, for that to be true, there needs to be a guarantee that these systems are trustworthy, can solve the problems they have been created for and can handle all edge cases within medical billing.
This growing interest in AI is reflected not only in market projections but also in its increasing adoption across health care operations. AI in the health care market is valued at $22.45 billion, with a projected annual growth rate of 36.4% from 2024 to 2030, according to a 2023 survey from JAMA. It is becoming a bridge between patients and providers, especially in medical billing departments. Doug Marcey, chief technology officer at Coronis Health, joined a recent Medical Economics Practice Academy session to unpack how AI is and isn’t transforming revenue cycle management (RCM), and why the future likely depends on a hybrid model of human and machine collaboration.
SOLUTIONS AND TAKEAWAYS
- Priority should be given to workforce augmentation over full automation to maintain a human connection between patients and providers.
- Invest in data governance and data science to make sure data within AI remains unbiased, instead of a waste of resources.
- Implement cost management strategies and continuous ROI monitoring to keep AI financially viable.
- Proactively address regulatory and security weak points to keep data safe and prevent legal consequences.
- Focus on specific, targeted AI applications with proven success for measurable outcomes.
DEFINING AI
AI operates under various models. Machine learning (ML), the broadest category, involves systems learning patterns from existing data to predict outcomes or classify information. This differs from traditional programming that relies on explicit rules. Many successful AI applications in RCM have roots in statistics, such as decision trees or regressions, which efficiently and cost-effectively process large data volumes. Deep learning, a more advanced model, mimics the human brain’s neural networks. It can learn new things and exhibit behavior that it has not been programmed to do. For example, deep learning will use pattern matching — based on voice or image recognition — to assign or classify data. When deep learning scales to massive proportions with billions or trillions of parameters, it leads to large language models (LLMs). These exhibit sophisticated, humanlike conversational abilities.
THE CURRENT AI LANDSCAPE IN MEDICAL BILLING
Assistance technologies have cropped up in several areas of medical billing. “AI has worked well when we’ve been able to design targeted models to do things like denial prediction and quality checks or work prioritization. This makes sure we have the work queue ordered the right way, so claims are worked on in an order that’s most cost-effective,” explained Marcey. Anomaly detection, like credit card fraud detection, is another valuable application in identifying billing irregularities.
“You can’t deal with the volume of data we’re dealing with in RCM today without some kind of automation,” Marcey pointed out. AI can assistin coding and claims follow-up by suggesting optimal ways to handle denials or rejections, thereby augmenting human capabilities rather than replacing them. For example, AI can present relevant ICD-10 codes, assisting a human who would otherwise need to recall from over 70,000 codes. LLMs can also power chatbots to provide patients with more accessible interfaces to their billing data, explain complex concepts and potentially reduce call center load. However, once patients are brought into an AI model, a line must be walked so as not to dehumanize the experience.
While promising, AI faces significant challenges in nuanced understanding, constant adaptation and robust data security. Autonomous coding struggles with the complexity of medical records and the vast number of similar diagnosis codes, making assistance models more effective than fully autonomous ones. AI models require continuous retraining to keep up with dynamic regulatory changes, payer policy shifts and evolving medical cases, which is an expensive task. “Just as we have our people constantly learning and improving, AI has to do that too,” said Macey.
Data privacy and security concerns are also paramount with large language models. They ingest vast amounts of data, posing risks of sensitive patient data extraction through prompt engineering that could result in a HIPAA violation. Laws like HIPAA were not designed with modern AI in mind, making it difficult to navigate issues around privacy, data usage and breach definitions. Integration challenges with legacy systems are common, too, as many electronic health record/electronic medical record systems are not designed to effectively feed ML systems. They sometimes lack direct key performance indicators (KPIs) or the ability to ban bots from accessing payer websites. To maintain data quality and remove biases, AI systems are entirely dependent on high-quality, clean, well-structured and unbiased data. “Most people are having to invest heavily in data science. These are specialists who pull out data sets and keep these problems in mind,” noted Marcey.
Lastly, return on investment (ROI) and cost management are a concern. AI initiatives incur increasing costs for training and running these models. Unexpected usage patterns, such as employees playing with a system or too much data upfront, can quickly inflate per-token billing for LLMs. “It can be easy to accidentally overendow the AI with more information than it needs, and that can really cost you later on,” warned Macey. Close monitoring of all associated costs is the way to help prevent this outcome.
HOW TO UTILIZE AI
To successfully integrate AI into RCM, practices and organizations should adopt a strategic, measured approach. The primary goal should be workforce augmentation, leveraging AI for repetitive, high-volume data tasks to free human staff for complex problems and critical decision-making. Implementing confidence intervals allows AI models to provide a confidence score with their predictions, enabling systems to escalate tasks with lower confidence scores to more experienced human agents for review.
Organizations should begin implementing AI with targeted solutions, defining KPIs for success, and continuously monitoring effectiveness as they scale up. That way, they can adjust or pivot if a solution is not yielding expected returns. “You really have to think of them as transformations of your business, making a small step, making sure you measure the effectiveness of that step, and then use that measurement to choose the next step and the next step,” said Macey. Involving human agents early, being transparent about goals and adjusting KPIs to reflect new workflows assists in building human trust and adoption. Given the complexity, consulting third-party experts can provide unbiased guidance and help navigate potential pitfalls without a vested interest in a specific vendor.
Ultimately, while AI is poised to transform health care billing, it is not a magic bullet. The future of RCM lies in a sophisticated collaboration between human expertise and intelligent automation. “Our people are the keys to making good decisions and being that trustworthy partner for our providers and our patients,” emphasized Macey. By managing data, following regulatory compliance, fostering human trust and diligently tracking ROI, health care organizations can harness AI’s true potential rather than succumb to its pitfalls.
Check out the full video and materials of this session.
