AI-powered computer vision, auto coder & denial brain.

The solution to manual and costly healthcare claims coding & denials.

The problem

Of an estimated $3 trillion in claims submitted by hospitals in 2016, an estimated 9% of charges ($262 billion) were initially denied. For the typical health system, as much as 3.3% of Net Patient Revenue, an average of $4.9 million per hospital, was put at risk due to these denials. Hospital administrative staff has grown at a whopping 3,200% between 1975 and 2010 in order to keep pace with the drastic changes in healthcare delivery, particularly change driven by technology and ever-more-complex regulations. These armies of administrators are doing little to relieve the documentation burden on clinicians, while creating layers of high-salaried bureaucratic bloat in the providers. The industry average for claims in A/R is approximately 63 days. In larger providers, A/R days can run as high as 97 days on average. Total days in A/R is creating a cash flow burden on providers. Additionally, of the $262 billion in claims originally denied in in 2016, only 63% was recovered, putting pressure on net income.

AI solution

In three months, introduced an MVP, that would help hospital executives overcome concerns with AI technology and embrace the opportunities of the solution to lower costs by automating the claims coding process, while at the same time dramatically improving accuracy leading to the overall reduction in claims denials.

Targeted results

70% reduction in claims denials 3% to 5% increase in net income 45% reduction in A/R days

In an effort to solve the growing risks to healthcare providers associated with extreme overhead driven by claims coding and claims denials, our client engaged us to envision, conceptualize, and build a white box, enterprise AI solution to automate the process and drastically reduce denied claims through a greater level of accuracy.

The business problem explained.

Today, healthcare providers are experiencing a much different set of risks – financial risks that can easily make the difference between making money or losing the business entirely.

Administrative jobs in healthcare continue to rise. The growth of these jobs combined with a shortage of talent has caused salaries to rise to new heights.

According to Health eCareers 2017-2018 Salary Guide, executive compensation has gone up 18 percent since 2016. That’s drawing a lot of new talent into the field, but experts say it’s a number that is unsustainable for creating long-term staffing solutions.

Health eCareers 2017 Healthcare Recruiting Trends Report notes that healthcare organizations are experiencing increasing employee turnover and a longer median time to hire replacement staff. Forty-one percent of healthcare workers reported that their pay increased over the previous year as a retention effort. And 64 percent of employers reported that they were increasing starting salaries in an effort to fix the gap.

Taking a deeper look, there is high demand for medical billers and coders. The Bureau of Labor Statistics (BLS) projects that the number of positions in the field of medical records and health information technicians — which includes medical billers and coders — will increase 13% between now and 2026. This industry’s growth rate is faster than the national average for all occupations, and approximately 27,800 new positions will be added by 2026. The dramatic increase in these positions are creating a recipe for disaster as costs are rising to unsustainable levels, continuing to put significant pressure on margins.

It should also be understood that medical coding is a 100% manual process in today’s medical provider practices. The manual processes along with the combinatorial math problem lead to human errors that perpetuate the claims denials; a problem that cannot be fixed with more human capital.

For instance, with the continuing drive to lower medical claims denials, there is a necessity to first understand why claims are denied in the first place. To do this, there is a need to perform a seven to ten factor analysis on every medical claim. With eighty data locations on a UB-04 claim form, over 122,000 medical codes with possibly 100 data features known about the specific claim (this is a low estimate), it can total up to 16 billion different combinations.

Denial Causes

AI solution — the future of work

Computer vision & auto coder.

The process.

Our goal, in partnership with our client, was to attack the boldest and most measurable problem(s) in provider healthcare.


Envision & Conceptualize.

In partnership with our client, we envisioned and conceptualized a process that would allow us to uncover the largest problems facing hospitals during admitting and billing. During this stage, it was clear that the coding a patient record presented a significant opportunity that resulted from combinatorial math problems that humans simply could not and cannot fix.


Make a plan

We then engaged in a full-scale development plan that would create a path to success through a series of MVPs (most viable products). MVP1 would be the most important phase that would determine the viability of the a prototype complete with:

Vision boards
UI & UX technological impact
Analytical modeling
Product roadmap
Business plan & go to market strategy
Detailed proformas
Data acquisition plan
Data governance and strategy plans
Working demonstration

Create a prototype created an MVP-1 working demonstration in 6 weeks that was comprised of a computer vision system & auto coder capable of meeting the hospitals where they are today. When completed, MVP-1 allowed a user to scan a PDF version of a patient EMR (electronic medical record) and receive an auto coded UB-04 form that could be dropped to the clearing house. Following MVP1, there are multiple MVPs that add features and improve the accuracy, speed, and overall learning capabilities of the AI.

The intelligent system of engagement explained.

AI computer vision

The computer vision system allows a hospital to upload a PDF of a patient record, then converting the free text and identifying procedures and diagnosis using the AI Computer Vision system. The computer vision system was created with the following features and functions in mind:

Speed of upload.
The ability to lift and interpret only the free text on the page. This means that any doctor handwriting, or imperfections as a result of scan or faxing needed to be addressed so that the free text could be read.
Accuracy of reading free text.
Interpretation of free text to interface with data dictionary.

MVP-1 was delivered with a 92% reading accuracy of patient records that had been faxed and scanned resulting in a less than 150dpi document. It is expected that the computer vision system will excel at 99% reading accuracy, processing 200 pages in under 45 seconds

AI auto coder

The AI auto coder is an intelligent system of engagement that provides a solution to the combinatorial math problem – capable of evaluating a claim based on a seven to ten factor analysis, resulting in over 6.1 billion combinations of codes and modifiers, and then applying the right ones for ICD-10, CPT, and HCPCS. These codes are then systemically applied to the correct claim form (e.g. 1500, UB04) that can be routed internally for other approvals or coding (e.g. charge master) or dropped directly to the clearing house for immediate processing to the payers.

The system can also read and evaluate an 834 form in order to assess and apply codes based on a patient’s policy coverage. The auto coder was initially trained with 50 million adjudicated claims and one of the largest data dictionaries in the industry. MVP-1 was delivered with 89% accuracy of coding.

As the product works through its MVP stages, the system will become smarter as a result of being trained with 4 to 6 billion records and is expected to reach 98% coding accuracy.

Denial Brain

The denial brain uses the power of high-quality data to determine the chance of denial of a specific claim and/or a group of claims following the coding process, but before the claim is sent to the clearing house. The ability to determine the likelihood of a claim being denied includes a seven to ten factor analysis of a specific claim and a comparison to similar claims in a group. Once grouped, the AI peers into the adjudication of all the claims in that group to figure out trends in the adjudication of each and every claim in that group – from the payer, to the type of coverage, to the insured, to the coding, to denials, and finally adjudication – and it all happens in less than 1 second.

But, the ability to get a chance of denial is more than a yes or no answer. With the chance of denial comes an explanation of why and also suggestions or things to ‘consider’ before the claim and bill is dropped. Perhaps there are changes that are needed to improve the cleanliness of the claim. Perhaps there is a blank or mismatched information on the claim form. Perhaps there is a modifier required. Perhaps the claim is missing documentation. Or in the best case, perhaps the claim and bill can be dropped with confidence.

The best part is with every claim processed; the denial brain is learning just like a human brain. Yet, unlike a human brain, it is able to consume the information from each claim, process it, and learn in real-time from billions of claims.