One of the things that has kept healthcare from becoming more convenient and cost-effective is the complexity of its processes. Human being are complex in every imaginable way possible. When a condition is diagnosed, there may be many aspects of the care that are delivered by different specialties and departments. All of this requires tight coordination and adds to the complexity of the whole process. It has been difficult to capture this complexity with traditional software in the last few decades. Software is meant to take over some activities for humans and automate things. In order to do that, it has to be able to capture a lot of this complexity. So far, that’s been mostly beyond the capabilities of the digital technologies available.
We talked about this in the last post in this series. If you want to relieve the care providers and the human coders (with this group, eliminating their jobs!) of doing manual coding, your technology needs to be sophisticated enough to read and understand the documentation of the care created by the provider. If you don’t do this well enough, you won’t get much adoption or will only be used for part of the process. While being able to assist humans and automating part of the process is good, it is not necessarily a game changer. You still need humans to do most of the work and review the output, so it’s progress but it doesn’t necessarily make a big dent in bringing healthcare in line with other sectors that have become far more convenient, like the travel industry.
Some of the breakthroughs in recent years in deep learning branch of AI may hold the key to change this. The launch of ChatGPT that has captivated everyone and the Large Language Models (LLM) that power ChatGPT are more than anything a breakthrough in the capabilities of natural language processing (NLP.) That means AI can understand written or spoken words much better, including clinical notes. Sure enough, we’re seeing companies that are launching AI products that are promising to take over some of these administrative tasks that are essential for healthcare providers but add more work to their already busy day. One example of that is ambient documentation. If AI can listen to the encounter between a provider and the patient and understand what happened and accurately document it, you’re eliminating a couple of hours of work for those providers. Similarly, if AI can read the document created from the encounter and generate the right codes, you’ve offloaded another tedious but essential activity from their workload. Given the shortage of staff and the burnout they’re experiencing, this could be a huge benefit to everyone in the healthcare ecosystem, including patients.
Autonomous medical coding is in its early stages but the results are proving to be very promising. Early experience with these solutions shows that they can achieve accuracy rates of 80%-plus, if not more. And, in a very positive development, after showing good proficiency for a period of time, the output does not have to be checked by human coders everytime. That means that the AI solution can read the note, understand what care was provided, decide what codes are appropriate, and perform the task of coding. This can then go directly to billing. That means offloading a whole lots of activities from providers, coders, and other administrative staff. If the bills are submitted directly after AI does the coding, this also significantly cuts down on the timelines involved. I discussed in an earlier post that getting paid in a timely manner is critical for medical organizations. If technology can do the coding and it’s accurate enough that in most cases it won’t need human oversight, then the timelines for submitting the bill and getting paid can be shortened. This is quite a benefit to these organizations and one that may actually motivate them to get their checkbooks out.
The experience to date for AI solutions in healthcare has been much buzz and hype but little in terms of clear results and documented ROI. There are many reasons for this, including lack of planning and funding to do real-world trials that show the impact of these technologies in the clinical and non-clinical parts of healthcare. Also many of the solutions developed were trained on limited data without the diversity of the populations they would be used for in the real-world. As such, the performance of many of these solutions has not lived up to the original promise. Another issue is that even if you have a good AI product, it has to fit within the existing information systems and workflows. The design of an AI-based solution that solves a problem end to end and simplifies existing workflows has been elusive for the most part. The AI coding solutions should be able to create the kind of ROI that is easy to document in a reasonable amount of time if their accuracy is indeed in line with the initial results from their deployment.
In the next post, we will examine how these systems deliver on their promise, how they evolved to their current state, and whether they’re ready for prime time.
