As the medical industry advances and as technology advances to help streamline operations for the business function of healthcare, AKASA believes it’s time to change what automation means in medical coding.
“In the classic Disney film Fantasia, Mickey Mouse, the sorcerer’s apprentice, asks a pair of mops to clean the place when they come to life,” wrote Varun Ganapathi, co-founder and Chief Technology Officer of AKASA. “But Mickey’s magic isn’t quite right, and the mops end up flooding the castle. In the real world, such automation failure can cost lives, especially in an industry like healthcare.” Ganapathi believes automation shouldn’t be static, it should work alongside human counterparts and learn from them to take on increasingly complex tasks over time.
To put this in context, at the recent Machine Learning for Healthcare Conference, Byung-Hak Kim, AI Technology Lead at AKASA told the story of Steve, a diabetic patient who had been admitted to the hospital by his primary care physician.
“After a few days of examinations and treatments, Steve went into the discharge procedure,” Kim said. “He received a discharge summary, where the clinical notes containing all the information about what precisely happened during his hospital stay is converted into a document using the standard format of medical codes.”
This summary, as Hak explained, is created by a professional medical coder, with the help of an automated clinical coding (ACC) engine, which leverages natural language processing (NLP) to spontaneously generate medical codes from clinical notes. Typically, the human coder will scan through the medical documentation found in the patient’s electronic health records, identify essential details underlying different treatments or services and annotate them with the ACC engine.
The job of human medical coders is becoming more and more complex as the primary system used to code medical bills (Current Procedural Terminology or CPT codes) has grown to include thousands of diagnosis and treatment codes that can be impossible to stay current on without the help of computers. Just one example of a CPT Code — Y93.D: V91.07XD: Burn due to water-skis on fire, subsequent encounter. Automating the translation of clinical notes in medical codes saves significant time and effort. However, it’s getting more difficult to accurately single out appropriate medical codes out of the given several thousands of high-dimensional options and literally “translate” unstructured free-text clinical notes into something that’s structured enough for administrative processes.
“Thus far, we have human beings doing the tough bit: Coders have to memorize every single plan and contract and thoroughly understand how they work to do their job properly,” Kim noted.“In the past three years, convolutional neural networks (CNN) and Long Short-Term Memory (LSTM) networks have emerged to tackle some of the challenges, but fully autonomous medical coding is still in its early days.”
Tapping into this opportunity and remembering the mistake Mickey Mouse made in Fantasia, AKASA was founded in the heart of Silicon Valley to design better automation for revenue cycle management in healthcare. AKASA’s technology combines AI and machine learning methodologies, human intelligence, judgment, and subject matter expertise to achieve robust and resilient automation.It adapts to the highly dynamic nature of the healthcare revenue cycle operations and assists healthcare systems to decrease their cost so that they can invest more in patient care and better steward every single dollar.
AKASA recently published peer-reviewed research on a novel AI model – Read, Attend, and Code – which is designed by employing attention mechanisms and transformers. This means that human coders can teach the Read, Attend, and Code model on the process of labeling and correct its mistakes before moving onto full automation. Recently, AKASA empirically tested this model in a first-of-its-kind study on the possibility of a novel machine-learning approach, which outperformed the present state-of-the-art models for automatic coding of inpatient clinical notes and professional human medical coders.
According to the press release, the AKASA team found that the Read, Attend, and Code approach considerably outperformed CNN and LSTM networks by 18.7% in Macro-F1 and human-level coding baseline. When comparing the RAC model’s performance with certified professional human coders, both were assigned the same coding tasks of more than 4,000 ICD-9 codes for a total of 508 inpatient discharge summaries randomly sub-sampled from the MIMIC-III testing set. This assignment took human coders about 30 hours over a week to complete.
“From this study, we’ve confirmed that the architecture we’ve designed to process clinical notes and automate annotating codes is now the state-of-the-art model for the industry,” Kim said. “By leveraging this model, we can streamline coding tasks to reduce excessive costs and coding errors while deploying resources more strategically to the most complex cases that need extra attention.”
Looking ahead, AKASA aims to maintain its research-first approach and make the discipline as transparent as possible to break through hype and provide easy-to-use AI tools that can be deployed into coders’ workflow. “Our study challenges others in the industry to be more transparent about how their solutions provide a return of investment to healthcare’s back office,” Ganapathi added. “This technology can help eliminate onerous paperwork that contributes to physician burnout and bloats healthcare spending.”