SubHero Banner
Social Share
Article In-page Navigation
Container for Section 1

Artificial intelligence (AI) is the foundation of data science. It's anything a computer can do that formerly was considered a job for a human. This is made possible by developing intelligent computer systems to think and work like humans. 

For example, AI can perform visual perception, speech recognition, decision-making and language translation. And this barely scratches the surface of what’s possible. 

AI has become such an active part of our day-to-day lives. We’re all familiar with using a virtual assistant such as Siri, Alexa or Google Assistant.


Machine learning and deep learning

Breakthroughs in AI, specifically machine learning, have further progressed what we can do. It’s driving much of the advancement we’re seeing in AI.

Yet another AI advancement worth noting is deep learning. The key to deep learning is its neural networks. Neural networks are algorithms that work like the human brain, recognizing relationships found in data. To train a network, we present it with large complex sets of inputs and millions of samples to “learn” the correct way to classify the sample.

Deep learning is transforming business operations. For example, property casualty insurers have made significant progress redefining manual work by using deep learning to review bills for accuracy. Rather than sample a small percentage of bills manually for inaccuracy, all bills can be scanned to find errors.

Container for Section 2

Deep learning is well-suited for health care

In health care, complex decisions often require skilled medical professionals like physicians and nurses. We can train neural networks to assist with these decisions using data from clinicians’ previous decisions about policy changes, prior authorizations and chart abstraction. There are opportunities to leverage AI in a variety of health care use cases to help medical professionals focus on more complex decisions and spend time with their patients.

  • Automating administrative tasks
    • Health care delivery such as prior authorization for pharmacy or special care, and recommendations for post-acute care
    • Population health management such as risk adjustment, Star Ratings, HEDIS® metrics, and recommendations for prevention and well-being
    • Health care operations such as auto claim adjudication, chart operations, network design, medical coding and provider directory accuracy
  • Helping coders focus on which charts to review Using AI to review medical charts is common for retrospective risk adjustment. We apply it throughout Optum chart review, which makes the process highly efficient.
    • Smart chart targeting: Predicts and prioritizes charts most likely to support specific unreported diagnosis codes
    • Digital retrieval: Helps identify retrieval modality deemed most likely to be successful with providers
    • Intelligent chart review:
      • Analyzes for potential unreported conditions
      • Routes reviews based on coder expertise
      • Improves automation of coder reviews
      • Detects if member’s health history may still indicate unreported diagnosis codes and routes for additional completeness reviews
  • Predicting disease risk Advanced data science methods and models can be used to support physicians. Using these models enhances clinical programs designed to identify conditions that could affect the health of their patients.

    Many electronic health records (EHRs) contain risk-adjustable conditions within the current problem list for a member. Examining an individual’s current health, health history, health ownership, social determinants among other factors can lead to a more accurate probability prediction.

    This creates a picture of a member’s disease likelihood or risk profile. And as it evolves over time, we can find out which clinical events were responsible for which changes.


Container for Section 3

A study in disease prediction

Background: We piloted a study for a health plan using disease likelihood analytics for 11 conditions across select provider groups. To score the likelihood of a condition, we used a disease likelihood analytic model trained on up to five years of clinical history. Newly identified members at risk for any of those 11 conditions along with a prediction score were delivered to providers so they could leverage the information at the point of care.


Key lessons learned: The prediction model was more effective than existing, more rules-based analytics at capturing unreported conditions and supporting providers with engaging their patients. This enabled the providers to report a more accurate and complete picture of member health.

The prediction score and tools took that a step further by helping providers focus on identifying conditions that might be in the early stages so members can be assessed.


Results: We saw an 8% increase in unreported conditions documented and an 11% increase of recaptured conditions compared to existing, more rules-based analytics. And even better, providers in the study reported higher engagement by using this data and tools.


Focused disease prediction: CKD
We discussed earlier how machine learning can predict disease risk much earlier. This enables health plans to intervene and providers to engage these members to help change the disease trajectory.

As we think about how to improve outcomes for members, chronic kidney disease (CKD) comes to mind. One in seven adults in the U.S. has CKD, which translates to about 37 million people.1 For that and the following other reasons, it is one of the first diseases we want to tackle.

  • As many as 9 in 10 adults with CKD do not know they have it1
  • Diabetes and high blood pressure are two main causes of CKD2 
  • Age, family history and race factor in
  • Heart disease is the major cause of death for all people with CKD
  • CKD is not risk adjustable until stage 4

The following is a basic sample framework that should be intuitive to how any of us could use data science to approach and intervene with an at-risk population and improve outcomes and risk adjustment accuracy.

  • Identify members with a high likelihood of CKD
  • Send at-home screening kits to those members
  • Apply National Kidney Foundation standards of care earlier to impact outcomes
Container for Section 4

Be selective about use cases

Excited to get started applying AI in your business and clinical workflows? Before you start writing use cases, consider these five areas of criteria regarding where AI makes most sense to help achieve your business goals.

1. Opportunity

  1. What is the opportunity to deploy the deep learning solution to the market?
  2. What is the core business value the opportunity would accomplish?
  3. Will the opportunity directly drive value, cost savings or an increase in customer and patient satisfaction?

2. Available type and amount of data

  1. What type of data is needed to address and train for the opportunity (e.g., claims data versus clinical data and/or medical charts)?
  2. Do you have a large amount of data to support the model(s) needed for the use case(s)?

3. Well-labeled data to train models

  1. What type of data is needed to address and train for the opportunity (e.g., claims data versus clinical data and/or medical charts)?
  2. Do you have a large amount of data to support the model(s) needed for the use case(s)?

4. Explainability

  1. What type of data is needed to address and train for the opportunity (e.g., claims data versus clinical data and/or medical charts)?
  2. Do you have a large amount of data to support the model(s) needed for the use case(s)?

5. Continue to train and maintain the models

  • Do you have the resources and skill sets we can dedicate to:

    • Track and evaluate experiment model performance
    • Select a champion model
    • Monitor data inputs and model inference
    • Retrain model
Container for Section 5

Four key takeaways

  1. AI is well-suited to health care because there are vast amounts of data containing decisions from skilled health care professionals. That data can be used to train neural networks to help automate the more administrative activities such as prior authorization in health care delivery, Star Ratings and HEDIS metrics in population health management and medical coding in health care operations. Using AI for those tasks may enable health care professionals to focus on more complex decisions and spend more time with their patients.

  2. Deep learning models and tools are showing promise in predicting disease and emerging risk for conditions for patients not previously diagnosed. Augmenting clinical decision-making with these models can illustrate the likelihood of illness or health events over time.

  3. Understand the data available to you against your use cases such as the amount, type, quality and labeling. Also, consider the “answers” you might use to train your models against. Prioritize processes for data governance, data management and data best practices.

  4. Be able to pull actionable insights from data. Organizations need resources who not only can perform strong data analysis but also continue to train and maintain your models to continually get the most from your AI business initiatives.


1. CDC. Chronic Kidney Disease in the United States, 2021. Last reviewed March 4, 2021. Accessed July 12, 2021..

 2. National Kidney Foundation. Chronic Kidney Disease (CKD) Symptoms and Causes. Last reviewed Feb. 15, 2017. Accessed July 12, 2021.




About the authors

Sanji Fernando, SVP, AI Product and Platforms, UnitedHealth Group
Sanji Fernando is responsible for the design and development of leading-edge AI models and analytics for the UnitedHealth Group enterprise, as well as the platforms needed to accelerate the development and deployment of these solutions.

Previously, Sanji led the UnitedHealth Group Research & Development Center for Applied Data Science (CADS). The CADS team applied breakthroughs in AI and machine learning to solve complex health care challenges for UnitedHealth Group by developing and deploying new software product concepts.


Eric Haseman, VP and General Manager, Optum Payer Solutions
Eric Haseman is responsible for end-to-end service delivery across Optum Payer Solutions including risk adjustment, quality, payment integrity and analytics. His focus is ensuring Optum delivers the results clients expect, advancing the business forward to address growing market needs, and achieving our mission of helping people live healthier lives and making the health system work better for everyone. For the past 20 years at UnitedHealth Group, Eric has continued to partner with clients on innovative approaches to achieve distinctive results together.


HEDIS® is a registered trademark of the National Committee for Quality Assurance (NCQA).

Tag: Health plans, Articles and blogs, Health care operations

Cross Promo Article(S)
Resource Library Side Bar

To learn more, watch the webinar on demand "How Advanced Data Science Can Predict Disease"

Learn more