Asset 1
Identify Members Prior to Major Medical Events

Unprecedented Data Presents a Target-Rich Environment  

As discussed in the first article, today’s environment provides payors access to unprecedented volumes and a variety of member data. Accountable care organizations, value-based contracts, payor-provider partnerships and payors’ vertical integration strategies all provide conduits for data exchange at an accelerated velocity with new healthcare information technology data communication standards such as Fast Healthcare Interoperability Resources (FHIR). The next step is leveraging that data for exponential value creation.  

Correct and Timely Identification – Harder than it Seems 

“Correct identification of persons at risk and the time when the model will be applied to inform intervention strategies is critical. Such identification requires a focus on the demographic characteristics, health status, and location of the patient population as well as on the clinical context in which a model will be used.”  
(Pencina et. al. New England Journal of Medicine, 2020)

Recent McKinsey research discovered that the biggest gap in payer success has been the correct identification of members for outreach and intervention — occurring only 20% of the time (Bestsennyy et. al. 2019).  This front-end limitation severely limits downstream effectiveness of the entire enterprise for value-based care (Figure 1, below).  

Figure 1. Opportunity for Improving Care Management Success via Patient Identification 

Traditionally, claims data is used to identify outliers and high-risk members. When this “hot spotting” technique was first applied, it led to a short period of some success – low hanging fruit. Value generation slowly eroded over time with diminishing returns on the strategy. 

Ultimately, using retrospective data to identify members with longstanding high utilization patterns for care management interventions achieves limited success because it’s difficult to decipher the driving factors and ability to change those factors (Williams 2015).

Artificial Intelligence (AI) models utilizing Natural Language Processing (NLP) for unstructured data are needed to revitalize payors care management and utilization management programs. AI & NLP will be game changers for payors in several key ways: 

  • Deep convolutional neural networks enable model building and analytic insights from unprecedented volume and variety of unstructured data; 
  • Machine learning discovers new relevant variables and causal factors unforeseen by humans; 
  • Probabilistic or Bayesian modeling transforms analytics into a prospective, predictive strategy where members can be identified before high utilization patterns start or high expense events occur; 
  • High velocity model development and configuration compresses the time from bench research to rapid deployment into clinical and administrative workflows; and
  • As target populations, the environment, new treatments or other factors change, machine learning models will tend to adjust and change in response — a dynamic approach rather than a rigid, rules-based static framework. 

All this comes together to mean something very simple for payors – but very profound. We’re on the cusp of a skyrocketing ability to accurately predict which members need which interventions at a point in time when something can still be done.    

Data Science: The Art of the Possible 

Reduce Unnecessary Care & Increase Appropriate Treatment

Overtreatment for low-value care costs the US healthcare system an estimated $75 to $100 billion annually (Shrank et. al. 2019).  Beyond costs, unnecessary treatment exposes members to potential harm. Cancer patients are especially susceptible to overtreatment (Welch & Fisher 2017) where the stakes are especially high, risk is difficult to communicate and clinicians frequently practice defensive medicine. For example, the machine learning approach to assessing high-risk breast lesions could have reduced unnecessary surgeries by 30% (Bahl et. al. 2017). 

Personalized Medicine through Machine Learning 

Individualized risk assessment models offer an opportunity to reduce overtreatment and ensure appropriate care. The wide variety of risk factors and logic variables involved in risk assessments such as NCCN guidelines make it impossible to manually apply and process to everyone in a member population. 

Using machines to accurately identify patients for screening, specialized treatments and predict disease risk presents one of the greatest opportunities for exponential value creation.  In the past few years alone, researchers applied convolutional neural networks, deep learning and machine learning techniques to prove this approach in nearly every medical specialty.  For example, machine learning models substantially outperformed the classic Framingham study’s clinical risk factors for hospitalizations from heart disease — Framingham predicted 56% accurately compared to the 82% from machine learning (Dai et. al. 2016; Paschalidis 2017). In oncology, nearly every cancer type has validated models applicable to risk assessment, screening and genetic variations   (Cuocolo 2020).

Table 1 shows the areas where machine-driven patient identification, treatment matching, and risk assessment are proven at some level and are being translated to live applications in payor populations. 

Genomics & Precision Medicine

Genetic testing for precision medicine is a challenging value equation for payors as the clinical utility remains uncertain for many diseases (Kogan et. al. 2018). There’s significant potential for machines to identify the members where a test poses high utility or even potentially replace the need for a test altogether. For example, radiogenomic machine learning models can automate genetic analysis by identifying visual signatures indicating the presence of genetic mutations which denote increased risk of cancer, cancer recurrence or other poor clinical outcomes (Li et. al. 2019, Cui et. al. 2020).   

Beyond Identification to Intervention 

Payors are in the best position to lead the charge applying AI & NLP in health care.  With the broadest view of the health system, payors can direct the accurately identified patients from AI & NLP models to the best care team member for appropriate interventions. In today’s world of patient-centered medical home (PCMH), accountable care organizations (ACO) and value-based contracting, that may not always be a physician.  For example, pharmacist-led interventions significantly improve clinical outcomes in chronic disease and show promise in economic and utilization metrics (Newman et. al. 2020).    

After accurately and rapidly identifying a member for an intervention, payor success depends heavily upon the strength and interconnectedness of the care management and utilization management programs. 

In the next article of our series, we’ll explore how AI & NLP serves as a universal knowledge platform supercharging payor care management and utilization management programs. 

Written By
Paul Clark

Paul Clark is our Director of Healthcare Research. Prior to joining Digital Reasoning he served as Vice President of Research & Education at The Health Management Academy, leading the research agenda and educational development for C-Suite executives at America’s leading healthcare systems.