Intelligent Clinical Coding: Rethinking hospital revenue (part II)
October 16, 2019
In my last blog, I described how ICD-10 codes are used inconsistently – a real problem if you’re trying to run a hospital based on your forecast revenue as a healthcare administrator, how the bottom line is that if you code more, quality suffers, and if you code better, productivity suffers. I mentioned how we’re talking about the accuracy of health payments totalling billions of dollars across the world, with $50bn being paid out annually based on coding in Australia alone. This time around, I’d like to address how the problem can be innovatively addressed by means of a human + machine solution created by an Accenture pilot project that has yielded positive results and might be the foundation for a future characterised by Intelligent Clinical Coding.
Human + Machine Solution
The non-innovative solution to this wicked problem is to recruit more clinical coders. Unsurprisingly, that approach has failed. It doesn’t break the Iron Triangle – so costs simply rise in proportion to the number of people being employed, insufficient coders are trained, remote and rural areas cannot access coders and the financial burden of health system administration increases, resulting in less funding going to direct patient care.
Intelligent Coding is an innovative approach that uses AI to get more output from the same level of input, without sacrificing quality, i.e. faster and smarter. It’s important to note that, in challenging the status quo, the fault does not lie with the clinical coding workforce. There simply aren’t enough clinical coders in Australia and the ones we have often haven’t completed all the necessary training.
Accenture undertook an Intelligent Coding Proof of Concept aimed at solving the challenge in line with the thinking of the Accenture Digital Health Tech Vision 2018 trend of Citizen AI. The trend is that AI “is much more than a technological tool—it is part of the workforce”, so ours is a human + machine solution. The coders need to continue the expert task of coding the complex case notes; where AI can play a role is in the coder-clinician interface. Currently, clinicians and clinical coders don’t communicate clearly on every episode of patient care. Coders simply capture information based on clinical terms they see in case notes. Even when there might be suspicions of poor documentation or other causes of under-coding, there just isn’t time to talk to busy clinicians whose primary focus is high quality patient care, not medical records and certainly not hospital revenue.
The Faster and Smarter solution
We pulled together a team of AI experts from around the world to look at both the speed and quality aspects of clinical coding and whether AI might be able to add value.
The first objective was to help the clinical coder to work faster by abstracting the relevant clinical terms from the medical records. This meant using Natural Language Processing to categorise and highlight those relevant clinical terms so that the coder did not have to read every word in the record to elicit the meaning.
The second objective was to help the clinical coder work smarter by predicting relevant ICD-10 diagnosis codes from the medical records. This meant using Convolutional Neural Networks (deep learning) to predict possible diagnosis codes so that the coder was prompted to consider additional diagnosis codes based on the medical record.
Finally, we had to build a simple, secure web tool that would allow the clinical coder to visualise the data in the context of their normal workflow.
Our Intelligent Coding proof of concept succeeded in predicting the diagnosis codes with a high level of accuracy. In 19 out of 22 diagnosis chapters, the combination of coder + AI was comparable to or better than a human coder alone.
It’s early days, this is really still a proof of concept. While the response from clinical coders in test environments has been positive so far, their main interest was in the efficiency that intelligent coding could offer, rather than in in the revenue assurance aspect that would interest hospital administrators. This outcome is probably natural, given that coders are typically measured on efficiency. The predicted diagnosis codes cannot always be assigned, the Australian Coding Standards must be adhered to and doctors may need to revise their documentation, however the potential for more accurate coding and better revenue assurance remains significant and vital. This is a critical step toward digital augmentation of clinical coders – cyber coders.
I am proud of our people who worked hard to get these outstanding results. I am proud of the health systems that enabled us to securely test the AI models on their de-identified data and finally, I have hope that we have moved a step closer to faster, smarter coding and a more productive health system that invests in frontline patient care.
I hope you’ve found this two-part blog stimulating. I’d welcome interaction if you have concerns, queries or comments, so feel free to reach out.