The Logic Behind Computer-Assisted Coding
By Susan Chapman
For The Record
Vol. 33 No. 2 P. 12
Used correctly, the technology offers organizations a chance to improve coding productivity with a few other goodies thrown in as well.
Used to review and analyze medical documentation and generate codes that align with the document’s contents, computer-assisted coding (CAC) can be compared to spellcheck in a word-processing program or the suggestions generated during texting. In all cases, the system “reads” the information and makes code suggestions based on those data.
As the health care industry prepared for the implementation of ICD-10, CAC became more attractive. “From a provider perspective, most organizations switched to CAC because of ICD-10 about five years ago. They felt the need to mitigate the challenges of adopting a new code set and gain efficiency,” explains Diana Ortiz, JD, RN, CDIP, CCDS, CCDS-O, marketing manager at 3M Health Care. “Without CAC, organizations would have struggled to keep up with the number of new ICD-10 codes and the volume of charts. Over time, the efficiencies gained from CAC have been applied to maintaining the quality of the documentation and coding.”
Katherine Lusk, MHSM, RHIA, FAHIMA, AHIMA’s 2021 president and chair, notes that ICD-10, coupled with the advent of the Prospective Payment System in Texas, a set-fee reimbursement system, impelled her former employer, Children's Health System of Texas, to adopt CAC. “Our coders would not have been able to keep up, and we decided to go to CAC to support them,” Lusk says.
Choosing the Right System
Most CAC tools are logic or knowledge based, both of which have limitations, according to Abboud Chaballout, founder and chief executive of Diagnoss, an artificial intelligence (AI) start-up. “Knowledge-based systems employ keyword search tools, which can be restrictive by nature, very stringent, especially when something falls outside of the parameters,” he explains. “Knowledge based is more about being able to provide a citation, providing justification for the codes that are being used. Essentially, the system says, ‘Here are the codes that we’re dealing with, and this is why this code is justified.’
“Logic-based software is rules based, and a knowledge-based system can be rules based as well. For instance, you have a procedure code that is defined by only one specific diagnosis code. Now that you have a match, here are the rules you can reference. Both of these types of systems assume that you know what you’re looking for. They basically take you on a journey to search for a code, and you have to know that you’re searching for a code in those instances.”
In response to the limitations Chaballout describes, Diagnoss has developed an AI deep learning system that acts as a “coding assistant” to a facility’s coders. “In a deep learning engine, you feed a machine samples of notes, like doctors’ notes, and then also the appropriate codes. Then, the machine begins to look at patterns, understand the data by deriving inferences from that data, and piece together the rules on its own,” Chaballout says. “In that way, the machine begins to understand what it’s never seen before. The codes it generates are not based on rules and keywords.”
Ortiz believes the AI approach is best. “But when you’re thinking of systems being built from an AI perspective, it’s crucial that they’re built on an expert system that includes a well-defined logic and accurate coding knowledge. Our CAC decision trees are based on an expert system built over many decades, which has been expanded to include both logic- and knowledge-based components,” she says.
An AI or a deep learning system can be facility specific because the technology adapts to the input and workflow. “There is too much variance in EHRs and the way we store data,” Chaballout says. “The generalized elements and facility-specific elements have to work in concert.”
Lusk says that logic- and knowledge-based systems have to be tailored to the facilities they support. “The knowledge has to be facility- and vendor-specific. None of the CAC vendors is an EHR vendor. Health care is really nuanced. Therefore, the CAC system has to be specific to the EHR vendor, the facility, and even the individual. All of the CAC vendors have to be targeted in those ways. They also have to operate within certain standards, and they do,” she explains.
Ortiz agrees that organizations need to be aligned with industry standards but also believes that defining what is considered specific for a particular organization is critical, noting, “While coding is standardized across the industry, organizations must decide what records and documentation will and will not be used to support coding. Deciding what constitutes your legal medical record and what documentation to feed your CAC system is a facility-specific decision. As an example, documentation queries may be part of the legal medical record for one health care organization but not for another. You want to be sure the CAC system is coding from the right documents, or portions of the right documents, as determined by your organization.”
The Ability to Learn
Deep learning systems are based on their ability to learn and adapt, but those skills are not as obvious with knowledge- and logic-based systems. For instance, Lusk says the system implemented by Children’s Health System of Texas back in 2012 has continued to adapt to both the coders and the environment. “Our coders got more efficient, and they leveraged the tools more,” she explains. “So, the system has learned. And, as time has gone on, the machine has continued to learn and grow.”
Ortiz concurs but adds a caveat. “CAC systems are always in learning mod, but there can be a big difference as to the scale at which the systems learn,” she says. “If a CAC system is applied to certain documents for just one organization, then it’s learning about only one facility. So, in that case, the system may not be learning enough to be effective. When a CAC system scales with a massive number of documents processed by the CAC engine across multiple facilities, then machine learning engages, precision and recall improve, and learning happens at a much faster pace.”
With regard to deep learning, Chaballout illustrates how the technology is able to adapt to specific organizations and coders. “The system isn’t waiting for the coder to search or know,” he says. “It’s constantly reading the note and priming the coder with suggestions. It’s kind of how advertising works online. You’re searching for a basketball online. The search engine now knows you’re searching for a basketball, and an ad is generated. All you have to do is click on the ad. In the case of deep learning CAC, a doctor documents a diabetic patient’s situation, and the system provides a number of appropriate codes, for example. Then, the provider or coder can select the right one, cross-check it against encoder tools if they want, and pick the best one. The system learns from each encounter.”
Chaballout explains that the system learns in the same way a self-driving car learns to navigate autonomously from point A to point B. Once the driver inputs an address, the car will determine the turns, speed, and other driving functions. If the driver knows a better route, then the individual is free to take over the steering wheel, and the self-driving car would retain that new knowledge.
“Deep learning CAC is much like that,” Chaballout says. “When you’re working with a knowledge-based tool, there is a gap between the documentation and the tool. Most current CAC systems rely heavily on the coder to come up with suggestions. AI-powered CAC, on the other hand, allows you to skip a few steps. Rather than making the coder come up with the search parameters, it’s making the suggestions autonomously based on the provider’s documentation, but it’s also open to input.”
Preparation and Going Live
Chaballout notes that prepping for a deep learning CAC system is somewhat different than the process would be for other types of systems. “Other systems require a lot of preparation and planning because you need to configure many custom rules and add multiple pieces of software, and people don’t always want to adapt to new programs. Our software is designed to require minimal behavioral change because it feels like it’s part of your existing EHR. You can leave your current workflow in place. We want to reduce the number of times/clicks it take for providers to pick codes. We want to improve the specificity of the codes picked and make sure to capture all the codes that could apply to what is in the provider’s notes. Our goal is to reduce the number of back-and-forth conversations between physicians and coders, reduce the amount of time it takes to generate a bill, and also keep the facility’s workflow as intact as possible,” he says.
Before getting to the implementation phase, Lusk and her team needed to demonstrate CAC’s importance to their organization’s decision makers. They engaged their executive and medical staff leadership so that they understood both the value and need for CAC. “We needed to sell it to them,” Lusk says. “The system had to accurately predict our population. We had to let the leadership know the value of it.”
Early on, Lusk and her team had discussions with the coders and clinical documentation improvement (CDI) specialists to assure them not only that their jobs would be secure but also that there would be opportunities to make their jobs more vital to the organization. “We let everyone know what they would get out of the process as a department and individually. We asked them what they wanted to see when the journey was over. We then could tailor the system and tell them how it would meet their needs,” Lusk recalls. “One of our goals was to increase productivity by 20%, and we were able to do that. We eliminated outsourcing, which was more threatening in terms of job security to the coders and CDI team than the machine was. We shared this type of information so that everyone would understand and feel supported.”
Consequently, the coders did not feel threatened. “We wanted them to put their best foot forward. With all these changes underway with ICD-10, we knew that couldn’t be possible,” Lusk says. “In order to serve the population, they had to understand the new system and have deep knowledge, something that required a lot of data. Our coders have integrity, and they were supportive of the change.”
Lusk says the implementation experience was smooth, thanks to the strategic planning and overall support that went into the process. “It was easy to go live because we were well prepared. After that first day, the coders were good, and [Children’s Health System of Texas] still has the same coders from 2012. They looked at it as, ‘This is a new world.’ They embraced it and felt they were making life better.”
James Norris, an inpatient coder for University of Texas Medical Branch Health, experienced a CAC implementation when his facility made the transition at the beginning of the pandemic. Norris was designated as one of two “super users,” individuals who were pretrained to assist others in learning the new system. “We took a pre-test to evaluate our level of knowledge and received the training first. After that, we met in person, before COVID, and had two or three days of training. We had some time to experiment with the new program. By the time we were close to the staffwide training, COVID hit, and we had to do it all online,” Norris says. “We did modules on our own, then had more training, and met online to discuss what we were learning and how the training was going. We’re all still adapting, especially since COVID meant we all began working from home while learning a new system.”
Ortiz believes the biggest challenge to adopting CAC is change management. “It’s a big shift. Coders are productivity driven and often compensated accordingly,” she says. “HIM management must recognize there is a learning curve with CAC and help coders prepare. Consider these questions: How can management get their teams ready? How can you help them understand there is a learning curve? What changes will need to be made to HIM policies and procedures? How will you monitor implementation?”
Ultimately, Ortiz sees CAC as another weapon in the coding toolkit; the technology simply makes coders more efficient and better at their roles. “The coder’s name is on the coded claim; they’re the final judge. Coders are under a lot of pressure to code accurately and quickly. That can be a hard balance to achieve,” she says. “To have a tool that suggests codes not only helps coders maintain accuracy but also helps them be more efficient. It elevates the role of the coder by automating routine coding, which allows coders to focus on more complex cases. And today, with the move to value-based care and growing quality reporting requirements, facility-based coders increasingly have to look at the same record in new and different ways. So, it empowers them to expand their skill set and develop their role into more than what it is today. CAC can help facilitate professional growth.”
Norris underscores Ortiz’s emphasis on accuracy, addressing the notion that some coders could become lackadaisical, believing the CAC system can do more of their job for them. “A coder’s QA [quality assurance] can come back to haunt them,” Norris says. “QA is very important to me, and you can’t always feel confident from what is generated. I’ve been coding for nearly 30 years and I check everything. As an example, COVID has symptoms like a fever, cough, and headache. But those can also be something else, something that isn’t COVID at all. You have to be careful that the system isn’t picking something up that isn’t accurate—like information generated from inaccurate documentation or documentation that should not be used.”
Some organizations may be overlooking helpful CAC features. For example, Lusk cites automated queries that can free the CDI team to work on more complex projects, prompts for providers when they’re ordering, identifying comorbid conditions that accurately depict a specific population—information that can be leveraged by case managers, and an automated assignment of diagnosis-related groups on admit day, including length of stay, to help managers plan for patient discharge.
“The system could also be leveraged in real time to identify potentially preventable readmits and complications while the patient is in the hospital,” Lusk says. “And now, with COVID, there can be augmented predictability, where we can train the system to identify things based on common words.”
According to Ortiz, there is always room for optimization. “HIM management may assume the coder is using the tools as designed and gaining in efficiency. But it’s important to challenge assumptions,” she explains. “We recommend reviewing the CAC reports and looking at the data to see whether guidelines and best practices are being followed. For example, are there individual coders who continue to code with codebooks in conjunction with CAC? There are always opportunities for improvement.”
Chaballout says organizations can measure the success of their CAC systems in two ways: observationally and statistically. “One way is for decision makers to simply observe how well the process is going. The other option is more statistical. Whichever method, or a combination of the two, the organization chooses really depends on its goals,” he explains. “If the goal is efficiency, we can help measure how much time is being saved getting an encounter done, based on reducing and easing navigation of the EHR. If the goal is accuracy, then we can measure the codes historically speaking, the way that they’ve been coded vs what could’ve been, based on the machine. Essentially, a retrospective study. It just depends on your goal and what questions you’re trying to answer.”
Ortiz believes that all organizations have to consider return on investment. “In a recent survey of customers, we learned that many organizations have been able to eliminate outsourced coding because their teams are working faster with CAC,” she says. “Equally important, the quality of the coding is getting better, a measure that all HIM teams want to achieve. When organizations are benchmarked against their peers, accuracy and quality are two important hallmarks of success.”
— Susan Chapman is a Los Angeles–based freelance writer and editor.