While finding and marking the right codes for a medical encounter or procedure seems like a trivial task in the big scheme of things, it is actually rather important. Allow me to elaborate. Not only those codes are used to submit charges for reimbursement, they are also used for monitoring public health issues, national statistics, national budgeting, and very importantly, they will be used to find the right groups of patients to train future medical AI models. That means inaccurate coding will have many consequences. You can misallocate resources and underinvest in key areas. You can choose data from a cohort of patients that doesn’t accurately represent the patients that will be managed using your AI model based on the fact that some of their charts includes the wrong codes.
All of this goes to communicate that while finding the right codes creates extra work for the healthcare professional beyond just documenting their encounter smack in the middle of doing their job (very annoying!) it is a vital part of not only how care is paid for but also planning resources and improving public health. As such, making sure these codes are as accurate as possible and taking out as much subjectivity as possible is a worthwhile effort. The codes chosen and the level of service coded can be very subjective. Also, since documentation and coding happens during a very busy clinical day, certain codes can be forgotten and never coded. This is human nature and historically, we’ve accepted this since we did not have technologies to help with it.
Over the last couple o decades, certain technologies to help with this have emerged. As the adoption of electronic health records has increased, it became possible to start thinking about using digital technologies that can help with choosing the right codes. Computer assisted coding (CAC) technologies found a place in augmenting humans with coding. Since care providers have relied on those providing the services to do the initial coding and then human coders that review the codes and the supporting documentation to ensure accuracy and the right level of coding for the services provided, technology can be used to assist these parties. This software has been used as another tool in the toolbox given the difficulty of of analyzing clinical narratives, which is how care is usually documented. The process of CAC begins with clinical documentation. The software analyzes the documentation using NLP algorithms to identify key terms and concepts in medical transcription. It then uses machine learning algorithms to suggest appropriate medical codes based on the identified terms and concepts. The suggested codes are then reviewed by a coding professional who makes any necessary adjustments or corrections. The final codes are then assigned to the patient record.
Until now (and maybe even now) the capabilities of the technologies available in accurately analyzing these notes and understanding the type and level of care provided has been limited. The algorithms used by these software used programmed logic, which only does a good job when it sees something it recognizes, or Natural Language Processing (NLP) that had limited capabilities given the sophistication of the technology available at the time. While CAC was first introduced in the 80’s, its widespread use was limited due to the fact that most care was still being documented on paper charts until the 2000’s.
Some argue that CAC improves accuracy by reducing human error, while others contend accuracy is highly dependent on the quality of the documentation being coded, as well as the time taken to train the CAC software. Simple and consistent phrases may lead to accuracy, but more complex cases still require human review. Either way, with the initial technologies used to train CAC software, the level of confidence in their output was not such that you would not have it reviewed by a human. Given that buying, implementing, maintaining software and training humans to use it in doing their job takes time and money, the benefit provided has to be high enough for medical organizations to justify buying and using it. For the traditional CAC software, this ROI analysis hasn’t always been clear and while its use increased over time, it never quite became a must-have for medical organizations. In fact, where it has seen the most uptake is with Revenue Cycle Management companies (RCM) who help medical organizations with coding and other revenue-related activities.
The two main implementation challenges are extensive system training requirements and high costs. Significant system training time is crucial. Ongoing training is needed to keep up with coding guidelines and changes to documentation practices. In addition, the purchase and maintenance costs of CAC can be prohibitive, especially for smaller organizations with limited budgets. While CAC aims to improve efficiency, realizing that potential requires significant investments of time, training, and finances. Organizations must weigh if those investments will pay off given CAC’s limitations with complex cases.
Ultimately, while CAC technology provides gains in coding productivity and efficiency, it has limitations in accuracy and coding complex documentation cases. As such, there is need for better technologies that can handle simple and complex medical coding tasks and their output can be trusted to be accurate and complete. Also, given the significant shortage of resources in healthcare, if modern technologies can take over some of these administrative activities like documentation and coding, those precious resources can be better used for the many other activities that need to be done in the course of healthcare delivery.
In the next posts, we will delve into the emerging role of AI in medical coding and how it can be a game changer in this area.
