New AI Approach Exceeds Performance of Traditional Models For Automatic Coding of Medical Notes

A new AKASA study finds the Read, Attend and Code machine learning model outperformed current state-of-the-art models for automatic coding of inpatient clinical notes by more than 18%

Researchers at AKASA^™, the only Unified Automation^™ company for healthcare revenue cycle management, published a new study highlighting how a novel machine learning approach outperformed current state-of-the-art models for automatic coding of inpatient clinical notes by more than 18%. The technology surpassed current models in terms of both efficiency and accuracy as well as the performance of professional human medical coders.

Read More: UNTUCKit Selects YOOBIC Digital Workplace To Power Store Operations And Associate Training

While most healthcare research typically focuses on highlighting innovation in a clinical setting, there is limited peer-reviewed research demonstrating the effectiveness of technology supporting the business function of healthcare. Of note, this is the first research that tracks output from machine learning models for autonomous coding and benchmarks to the performance of human coders.

“We’ve reached a new milestone in autonomous medical coding technology as Read, Attend and Code (RAC) iscn the new state-of-the-art model in this space,” said Varun Ganapathi, Ph.D., co-founder and CTO of AKASA and co-author of the study. “Our publication challenges others in the industry to be more transparent about how their solutions provide ROI to healthcare’s back office. This technology can help eliminate onerous paperwork that contributes to physician burnout and bloats healthcare spending.”

Researchers compared AKASA’s RAC approach, which is designed by employing attention mechanism and transformers, to the performance of published state-of-the-art machine learning methods based on Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks as well as the output of certified professional human coders.

RAC considerably outperformed the prior art based on CNN and LSTM networks by 18.7% in Macro-F1 and also surpassed the human-level coding baseline. When comparing the RAC model’s performance with certified professional human coders, the RAC model and human coders 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,” said Byung-Hak Kim, Ph.D., AI Technology Lead at AKASA and co-author of the study. “By leveraging this model, teams 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.” 

Read More: Code42 Incydr Detects Corporate Data Exfiltration Movement To Personal Google Drive, Box And Other…