Episource Boosts Productivity, Delivers Quicker Healthcare Insights Using Machine Learning

Episource works with healthcare organizations and health plan providers to manage their risk adjustment and quality programs through data analytics. One of Episource’s main offerings is analyzing clinical documents to deliver insights that help clients estimate future healthcare costs, calculate their patients’ health risk scores, and segment their member bases for better planning.

“Scalability is our primary concern as a company. We interpret vast amounts of data to measure the performance of our clients’ data, analyze their member populations, and provide insights on their risk adjustment programs. It’s not a matter of if, but when we will need to process more data, and as such the question we always ask is: how can we make our platform more scalable?” says Manas Ranjan Kar, vice president of data science and natural language processing (NLP) at Episource.

With Amazon Web Services (AWS), Episource built a scalable information extraction pipeline to improve the processing of its clinical documents. In 2019 alone, it handled over 100 million pages of customers’ clinical documents.

The digitization of health care—along with the advent of electronic medical records (EMRs)—has created a deluge of information spanning patient records, to clinical data, to payer and provider statistics.

Kar shares, “Most of the information extraction and analysis of unstructured healthcare data has typically been manual and resource-intensive. Usually, these operations are carried out by domain experts who would have undergone a fair amount of training and certification to carry out the tasks at a ‘gold standard’ of accuracy.”

He continues, “Risk adjustment and clinical information extraction are highly knowledge-dependent. It would take multiple training programs across 6-7 months for a new hire to achieve a proficiency of 95% accuracy at processing clinical documents for our clients, which is the standard that we hold ourselves to. Our executive leadership had a vision for this process to be streamlined and the company to see renewed growth, without being held back by the constraints which usually accompany any knowledge-intensive operation. To achieve this vision, we actively started looking at how we could integrate machine learning (ML) and NLP to build a more scalable platform.”

Episource knew that ML had the potential to improve the process of clinical information extraction through the use of NLP and cognitive algorithms. ML could help analysts interpret unstructured text and process specific data points in clinical documents more efficiently. It would not only aid new hires but also senior analysts—who could trust the platform to reduce any repetitive tasks, while ensuring that quality levels were improved across the board.

The team began building its own machine learning platform with open-source tools, such as Keras, PyTorch, and Scikit. However, there were unique challenges pursuing this path—getting disparate tools to work seamlessly together in the development cycle often resulted in breakdowns mid-process, and the engineering team spent time troubleshooting intermittent errors on the ML platforms.

Episource sought to streamline the development, testing and deployment of ML models that it could use to accelerate the processing of clinical documents. It chose AWS for its seamless integration with open-source tools, the cornerstone of Episource’s ML platform. Episource leveraged integrated ML solutions, like Amazon SageMaker, Amazon Elastic Container Registry (Amazon ECR), and Amazon Elastic Container Service (Amazon ECS), that provided it with a more comprehensive overview of its ML platform, reducing exceptions and helping it to manage open-source ML tools more effectively.

“ML is a quickly evolving field, and it has increasingly turned engineering-heavy as well. This is why compatibility of our ML Ops pipelines with open-source tools is such an important factor for us—you cannot continuously build on your ML offerings without open source. AWS provides us with a setup that integrates seamlessly with open-source solutions, and allows us to leverage it for our benefit,” says Kar.

By using these solutions, Episource was able to deploy ML models to production quicker, and more cost-effectively. Its ML model development cycle was reduced from months to days. The engineering efforts focused specifically on automation, scalable architecture and tightly knit pipelines to allow Episource’s data scientists to experiment at will with development. For example, the engineering team leveraged multiple AWS services to build distributed and serverless data processing pipelines, CI/CD for ML, packaging of ML models and libraries, Infrastructure as a Code and custom ML workbenches. With ML algorithms simplifying the document review process for staff, the quality levels of the clinical information extraction process increased as well.

The integration of AWS’ services—such as Amazon SageMaker, Amazon ECS, Amazon ECR, Amazon Elastic Kubernetes Service, AWS Batch, and AWS Glue—with open-source technology has allowed Episource to build a ML platform that is not only scalable but also adheres to best practices of data security and privacy norms for healthcare. Its developers and data scientists are now empowered to build customized, innovative and multi-modal ML solutions that work best for its current, and future needs.

“AWS has provided us with a solid foundation to build up the scalability and efficiency of our ML products. The nature of our business means that we will always have tons of data to process, and more insights to deliver. AWS is a platform that takes continuous innovation as seriously as we do, and we are excited to see how we can keep working together to deliver better outcomes to our customers,” adds Kar.