Optum Focuses On AI To Improve Administrative Decisions

Optum is a giant within a giant. It’s a business unit within UnitedHealth Group, which is ranked 7th in the Fortune 500 and had $242B in 2019 revenues. Optum itself had $113B in 2019 revenues, and is focused on three areas of health care: direct healthcare services, pharmacy services, and information- and technology-based services to many healthcare providers and payers, including United Healthcare.

By bringing together a history and culture focused on innovation with industry expertise, diverse data sets, advanced analytics and emerging technology, Optum works to propel health care forward in a way that better serves everyone.

Data Science Applied to Healthcare

OptumLabs was founded in in 2013 as a healthcare innovation and research center. It collaborates with partners from across the healthcare ecosystem to improve patient care and patient value. Data science was a strong focus of the Labs from the beginning, focused on extracting insights from both Optum datasets and those from partners. As is common in healthcare data science, data scientists at the Labs were interested in improving clinical decision-making. Researchers there have analyzed extensive data and published widely on treatment strategies for many different diseases, including maternal health, Alzheimer’s disease, diabetes, heart failure, opioid addiction and COPD. 

Somewhat less commonly, however, the Labs also began to address the potential for data science and AI to shed light on administrative processes, like revenue cycle management—how healthcare providers get paid for their services. That set of issues became the primary focus for Sanji Fernando, who started at Optum Labs as VP of Innovation in 2014, and eventually became the head of the Labs’ Center for Applied Data Science.

Fernando is passionate about the opportunity to use artificial intelligence (AI) and machine learning (ML) to improve administrative processes in healthcare:

We studied some of the recent breakthroughs in data science and AI, and how they might be applied in healthcare at the Labs. We put more attention and investment into AI breakthroughs like deep learning. We realized that they could be very helpful in improving productivity in key areas like revenue cycle management. Improving reimbursements and coding medical charts are tasks well-suited to neural networks. This became one of the key hypotheses while I worked in OptumLabs. While there is great opportunity to use machine learning for clinical decision support, there is just as much opportunity to optimize administrative processes. There is no doubt that AI will transform clinical care, but the lower hanging fruit is administrative challenges.

In particular, Fernando believes that ML can facilitate the process of “getting to yes” across the payer/provider interface. He noted, “We can use the technology to find where both sides agree, and quickly resolve those transactions. People then can focus on the hardest and most complex cases.”

As the potential for this set of use cases became apparent, Fernando moved from OptumLabs to the Optum Enterprise Analytics organization, where he is Senior Vice President of Artificial Intelligence and Analytics Platforms. Instead of researching the role of AI in administrative processes at Optum, his role now is to put it into production.

A Use Case in Claims Reviews

Optum employs a number of physicians who provide third party review of provider claims; the review is required by many payers before they will pay the claim. At Optum and elsewhere there is a tremendous work queue of clinical cases to be reviewed. Whether or not the case supports an inpatient admission to a hospital is a key issue in the review, with substantial revenue cycle impacts.

Fernando and his colleagues thought about what they could do to enable that process so that the easy cases are automated, and the Optum physicians are only reviewing the hard cases. They knew the power of deep learning to correctly classify text and used previous decisions of Optum physicians to train neural network models to read medical notes and determine whether the case supported an inpatient stay and other types of care.

Once they had a successful model, Optum worked with provider customers to evaluate all cases, providing each a score that suggests how likely the case is to be approved. The providers themselves could then decide what cases need review. Those that seemed likely to be well supported for an inpatient admission could then be examined by the human physicians, allowing physician reviewers to focus on the most complex cases.

Based on this first success, Fernando’s team learned a great deal about the challenges to develop and maintain ML models, especially deep learning models. The team then embarked on developing a platform to deploy the models, manage data pipelines, and monitor the accuracy of the decisions over time. If they found “drift” in the models—suggesting change in the features of the input data—they knew they needed to be retrained. The team can select among several different neural network architectures with different numbers of levels to find the one that best fits the data for any given training dataset.

From Data Science Research to Deployment

Having led innovation functions at Nokia and OptumLabs, Sanji Fernando has a clear notion of the distinction between research and innovation. Research can be done within a lab, but innovation, he says, requires building trust and relationships inside the operating parts of the business. The business leaders, of course, understand the business better and become partners in innovation. “We have sometimes thrown breakthroughs over the fence at people,” he admits, “but they rarely get implemented.”

He also argues that a key aspect of innovation is admitting that something doesn’t work. “It’s hard to admit that a great idea just did not work, and we should walk away from it.” Being candid with the business can build a great deal of trust with internal business partners and lead to future success.

Optum’s business involves millions of transactions, many of which benefit from more automation, data-based decisions and speed. The work of his group in applying AI and ML to frequent transactions, limiting human involvement to the situations in which it is truly necessary, is having a major impact on the quality and cost of healthcare.  

AI and Future Potential

AI can be used to solve a variety of health care challenges, from improving recommended care pathways, aiding decision making, reducing friction and delays in health care processes, and improving clinician productivity—enabling them to spend more time with patients. Given AI’s potential, Sanji Fernando believes it will play a major role in shaping the health care system to operate more seamlessly to support preventative, holistic patient care.