Building the Autonomous Coding Ecosystem: What HIM Leaders Need to Know

The United States is the biggest medical coding market in the world and will continue to grow, nearly doubling in market size between 2020 and 2028, according to a recent CRISIL GR&RS Analysis.¹ Hospitalizations, surgeries, and case complexity are also on the rise, contributing to higher coding volumes. McKinsey & Company predicts inpatient admissions and operation room procedures in 2022 will be 4 percent higher than in 2019 according to its late 2021 survey. Furthermore, the CRISIL GR&RS Analysis also predicts that the outpatient segment is anticipated to grow 9.8 percent by 2028.²

Amid this ongoing expansion of healthcare encounters, a confluence of challenges faces health information management (HIM) and revenue cycle leaders. Pandemic-led turbulence impacted the availability of qualified coding professionals as senior professionals retired, medical leaves occurred, and other coding professionals shifted careers.

The need for faster and higher quality coding outcomes to ensure revenue integrity, reduce reimbursement recoupments, and bolster bottom lines adds pressure to these coding challenges as healthcare provider organizations emerge from two years of unpredictable cash flows. The average 350-bed hospital revenue cycle leaves $22 million on the table (an estimated 3-5 percent of revenue).

Quality coding is essential to stop this type of revenue leakage and keep healthcare provider organizations in business. Coding is the most human capital intense part of the whole revenue cycle, but also the most critical, as it relates to reimbursement outcomes. As technology changes, HIM and revenue cycle professionals must embrace innovation as the most practical pathway to address these challenges.

Navigating the Perfect Storm

While coding technologies have evolved over the past 35 years, many have fallen short of productivity expectations. Technological gaps, along with the “Great Resignation” and ongoing coding professional shortages, have created a perfect storm for the revenue health of provider organizations.

Tools like encoders and computer-assisted coding (CAC) simply augmented coding professionals’ clinical knowledge and extensive coding experience. They didn’t truly automate coding processes and are not always able to accurately determine when cases required human intervention. This is where intelligent autonomous coding differs from CAC and its predecessors.

Autonomous coding engines use sophisticated, new generation computer algorithms to code charts within seconds, with zero human intervention and proven high accuracy. The intelligent autonomous engines understand relevancy and which codes are most accurate to assign.

If the engine cannot interpret the documentation and translate it into ICD-10 codes, CPT/HCPCS codes, E&M levels, or modifiers with a high level of confidence, it is considered an exception, and the case flows to a human coding professional for traditional coding. Items such as missing documents, ambiguous sentences, complex CPT codes, obscure diagnoses, and client-level business rules are examples of what may trigger an autonomous coding exception.

The Technology Within Autonomous Coding

There are a few different scientific approaches to building an autonomous coding engine that raised the bar of success over the last few years, including computational linguistic modeling and machine learning. However, the foundational technology is consistent and includes technology coding professionals have been using for several years, even if they don’t realize it.

This foundational technology focuses on the linguistics, breaking down the sentences and mapping to machine language. The advances in autonomous coding engines are mainly a result of next generation machine learning, which understands past and present to predict the future. Patterns are identified and learn over time to improve outcomes. From there, another level of pattern technology is applied: deep learning.

Deep learning is part of a broader family of artificial intelligence tools. It uses data patterns to predict the future. Autonomous coding’s deep learning patterns are built in the same way that Amazon’s Alexa builds your favorite playlist, Siri recognizes your preferences and makes related recommendations, or cars self-park and read your GPS location.

The difference with deep learning is that rather than a lot of rules being constantly written and updated, the system learns from the data and continuously creates new rules. Therefore, autonomous coding is very successful in high-volume outpatient service areas such as emergency department and radiology.

The system requires a lot of data, but it infers new rules and unique variations as it continually adapts to the information it receives. It’s important to understand what technology elements are needed for a successful autonomous coding engine.

The Four Technology Layers Needed for Platforms to Succeed

For machine learning, deep learning, and computational linguistic modeling to work effectively, the paragraphs, sentences, and words must be broken down correctly. Natural language processing (NLP) plays an essential role since the system must first be able to communicate via human language (“natural”). NLP is a foundational technology also used in CAC and computer assisted physician documentation (CAPD).

NLP accuracy has dramatically improved since 2017. Formerly based on decision trees, NLP today employs statistical algorithms to find relations, dependencies, and even context. For example, there must be a machine translation of the sentence “Heart attack treated with Bayer” to discern between aspirin and a grizzly. Standard terminology systems such as the Systematized Nomenclature of Medicine, Clinical Terms (SNOMED CT) aid NLP by providing a computer processable collection of medical terms and mappings which are also the lexicon foundation for most electronic health records (EHRs).

Layered on top of NLP is natural language understanding (NLU). This is a subset of NLP that interprets language, derives meaning, identifies context, and draws insights. For example, in the sentence “Please crack the windows, the car is getting hot,” NLP focuses on the literal sense (picture an actual cracked window), while NLU extracts the context and the intent (draws inferences to determine what is meant).

Another technology tool embedded within autonomous coding is clinical language understanding (CLU). This is an upper layer of clinical intelligence driven by subject matter experts. Doctors, nurses, and other clinicians interact with data scientists, IT developers to teach the system clinical aspects of each case and pass the “human test.”

The final layer for autonomous coding success is the application of business rules. These encompass rules that focus on the coding guidelines, payer rules, edits, client specific business rules, etc.

Autonomous coding vendors take all four layers into account when developing their engines. Machines learn where to find the information they need to normalize the context of language for coding. Factors such as high volumes (needed for machine learning), small document sets, lower case complexity, structure of documents, and code set specificity are also considered during system development.

The Demonstrable Results of Autonomous Coding

There are a handful of vendors showing demonstrable results with autonomous coding. Typical offerings include facility and professional fee coding for radiology, emergency, urgent care, hospitalist medicine, laboratory, pathology, primary care, and other high volume outpatient encounters. Since the technology is so complex, vendors tend to specialize in these very niche outpatient service types.

Future case types for autonomous coding include gastrointestinal, women’s health, vaccines, cardiac rehab, outpatient therapies, and more. These clinical specialties are prime candidates to have their encounters coded more effectively through advanced autonomous coding platforms.

However, juggling multiple autonomous coding vendors to support niche areas is rarely a viable option for hospitals and health systems. Therefore, a best of breed platform approach has become the optimal way to implement autonomous coding, while also respecting IT staffing burdens, coding professional staffing efficiency, and capital costs.

Platform vendors vet the best niche autonomous coding engines and integrate them into a single system. The vendor evaluates, integrates, and manages each application and presents them as a single, united solution. For example, the platform intelligently routes cases to the optimal engine for coding based on the patient type (e.g., radiology, emergency department, urgent care, hospitalist, etc.) using business rules.

The Impacted of Autonomous Coding

The three biggest questions I hear from healthcare leaders are 1) “What is the lift for the IT team?”; 2) “What is the commitment for implementation?”; and 3) “What does this mean for  coding professionals?”

For IT teams, the most important precursor to autonomous coding is confirming the data integration approach. This upfront initiative is owned by the autonomous coding vendor, as they have mastered ADT/document data exchange, EHR/billing system integrations, and clinical document parsing. They also will conform to the organization’s preferred approach resulting in minimal effort for the organization’s IT team.

Implementation efforts vary by the autonomous vendor you select. There are typically five common steps that are jointly owned by the autonomous coding vendor and the healthcare organization. Implementation averages 90 days but can be fast-tracked with the engine coding cases as quickly as within one month. The facility works collaboratively with the vendor to customize based on their unique needs. Here are five common implementation steps.

• Discovery: Vendor analyzes all current workflows and internal guidelines. Custom profiles are created for user rules engine (some allow customers to create on their own and then they approve).
• Engine Training: Vendor processes a controlled amount of data to train the engine and reviews output for quality. This step is performed repeatedly until the engine achieves at least a 95 percent accuracy score or higher. This step is not needed for the computational linguistics model because deep learning and pattern prediction is not a part of this autonomous technology approach.
• Integration: Vendor customizes integration points for the client (HL7, secure file transfer, CSV, API, etc.).
• Shadow Period: Coding professionals code in parallel with the machine. One hundred percent of the machine output is compared to prove out quality prior to live production. Proof of concept is completed here. Not every client requires this step and the time of shadow can vary.
• Go-Live: Continuous tuning at go-live and following go-live ensures automation rates increase over time. Also, quality control is aligned to ensure a continuous 95 percent+ accuracy rate.

How to Measure Success

Measuring success is an ongoing responsibility for any technology implementation. This is especially true for autonomous coding. Most vendors provide data analytics and dashboards with regular measurements of success criteria.

Metrics such as percent automated, percent human coded, time to bill, and quality should be included in the analytics by specialty. You will typically experience 70 to 98 percent of cases automatically coded by the machine with a 95 percent or higher accuracy rate depending on service line. If the system is not confident in a correct code assignment, the case is automatically routed to a human coding professional, as mentioned above. Based on industry standards, you should not expect anything less than 95 percent accuracy from autonomous coding—just as you would expect from human coding professionals.

In addition to these results, the ability to scale without adding manpower, continuous improvement in automation rate, and a 15-20 percent reduction in coding expenses are typical key performance indicators for autonomous coding implementations.

What to Expect Next

Autonomous coding has come a long way. The top vendors continuously improve their engines with the latest data science and technology. The engines are getting smarter, the type of visits are continuously expanding, and the automation throughput is quickening with high-quality results.

Vendors’ focus is also narrowing. Therefore, to optimize autonomous coding across outpatient areas, HIM leaders may need to partner with multiple vendors.

Finally, most vendors do not have a coding platform to accommodate coding workflow for the coding of exceptions. Revenue cycle and HIM leaders ideally should keep a pulse on current technologies available, how these technologies can impact various service lines, and prepare organizations for change based on that understanding.

Examples of Success

The return on investment for autonomous coding can be impressive. For example, a free-standing radiology center went live with autonomous coding in 2020. At this time, the coding professionals’ average salaries were $55,000 per year, but, fully loaded, the cost per coding professional was $73,150 annually. The center had a million encounters yearly, and human coding professionals averaged 20 cases per hour across 28 employees, resulting in annual coding expenses over $2 million. Autonomous coding technology dropped the cost to code per case (CCPC) for this center to $650,000 annually, a savings of over $1.4 million.

In a second example, autonomous coding was implemented at a university hospital for emergency department and injection and infusion visits. The organization’s average annual coding professional salary was $60,000 per year, but, fully loaded, the cost per coding professional averaged $79,800 per year. Volume and productivity were lower in this example, but the organization still experienced a net savings of $938,298 per year.

Other results for these organizations included decreased DNFB, lower coding days, and acceleration in cash collections.

Shifting to Human-Assisted Coding

Autonomous coding is human-assisted coding. It optimizes all the recent technological advancement in machine learning, deep learning, NLP, and clinical language understanding to propel coding professionals beyond the disappointments of CAC. Humans are needed to manage exceptions only, not every case.

Autonomous coding frees up coding professionals to code more complex cases while relieving the staffing, expense, and productivity burdens of the routine.

Notes

1. Market Analysis US Medical Coding Service: CRISIL Global Research & Risk Solutions. September 2021

2. Ibid.

Monica DuBois (monica.dubois@deliverhealth.com) is the vice president of coding solutions at DeliverHealth.

Learn about the innovations and technology in healthcare that will impact your work and your career by participating in the AHIMA Evolving Healthcare Webinar Series, which includes the on-demand webinar “Artificial Intelligence in Healthcare.”