Ask Dr. Parker: Predictive Analytics in Care Management
In his latest advice column for leaders in value based care, Rich Parker, MD explains why care management is so valuable – and how predictive analytics can help identify the patients most likely to benefit.
“My ACO has limited care management resources. How can I make sure they go toward the right patients?”
Those of us who practiced medicine before care management came into being can easily understand why it has become an integral part of most ACOs: It’s better patient care! It is easier for a doctor to work as part of a team dealing with complex and challenging patients instead of trying (and often failing) to do it alone.
I want to share some of the science and art of one aspect of care management – namely, how to identify “actionable patients”. The sickest patients, who are relatively easy to identify, are not necessarily the most actionable. For example, the patient with end stage renal disease on dialysis or metastatic cancer may be very ill, but not particularly amenable to care management interventions.
An overview of care management
But let’s back up for a moment. Care management usually describes a set of functions carried out by a registered nurse trained to handle a cohort between 50 to 200 patients. Commonly, the model includes telephonic contact between the nurse and the patient, with occasional in person visits either at the doctor’s office, at the patient’s home or even at some other location such as a local Dunkin’!
Typical care-managed patients have several illnesses, long or complex medication regimens, and physical or socioeconomic challenges, often including behavioral health issues. Before care management, these were the very patients who often could not get to the doctor’s office, or did not refill medications, or missed out on physical therapy and other needed services. This frequently resulted in unnecessary emergency department visits, hospitalizations, and other potentially avoidable high cost utilization of healthcare resources.
Care managers establish good rapport with their patients and have an excellent track record of improving the health of the patients. This results in fewer emergency department visits, fewer admissions and lower overall medical utilization.
Managing limited care management resources
The holy grail of care management has been to find the elements predictive of future high utilization and amenable to improvement by care management intervention. In the real world of the ACO, resources are limited and must be allocated across multiple legitimate competing needs. A medical director or chief medical officer needs to make a compelling case that expending resources on care management will result in both better health outcomes and substantial decreased utilization net of the cost of the program.
Understandably, care management programs initially focused on identifying “the sickest” patients. This approach resulted in a mix of manageable and unmanageable patients enrolled in care management programs – a suboptimal use of resources.
Care management programs have attempted to use a number of existing algorithmic models to yield a higher percentage of actionable patients. These are traditional morbidity-based models, focusing on current disease burden, and include the CMS’ HCC and Johns Hopkins’ ACG, among others. However, these models have limitations. They are not able to incorporate some critical elements for determining whether a patient is impactable, such as socioeconomic factors or near real-time physiological data such as BMI and blood pressure.
An alternative: finding actionable patients with predictive analytics
Predictive analytics is the science of choosing the correct elements with the correct weighting resulting in the most accurate guide to any individual patient’s future utilization, usually over a 12 to 24 month period. Predictive analytics offers an alternative to traditional risk adjustment methodologies in finding patients most likely to benefit from care management.
My colleagues have worked hard to address the question “What factors best predict which individuals will be most benefited by care management?” where benefits are measured in terms of reducing medical events (utilization). Our goal was to improve patient quality of life while reducing the cost of care.
To do this, we identified a large, diverse sample of individuals (broad demographic and medical profiles) covered under a range of commercial, Medicare, and marketplace payer arrangements. Within this group, we identified a smaller subsample of individuals who had been enrolled in a care management program and had at least one interaction with a care manager. In both cohorts – care managed and “unmanaged” – we identified total cost of care and utilization 12 months before and after an index date.
To model cost and utilization for both cohorts, we used both utilization data throughout the sampling period as well as a number of other factors, including morbidity risk, care coordination, frailty concepts, and aggregated diagnosis groups (per Johns Hopkins ACG), census variables (such as vacancy rate and high school graduation rate), clinical observations (such as blood pressure and BMI), and other utilization statistics.
Our tests of several applicable models led us to use a logistic regression for cost and avoidable utilization events, with the difference between these two populations representing the “impactability” of an individual.
This impactability is an estimator of the reduction in cost and utilization for patients enrolled in care management relative to unmanaged patients. Using predictive modeling, we were better able to model and define risk stratification compared to typical risk adjustment methodologies.
By projecting utilization, cost, and patient outcomes with and without care management, we can identify those for whom care management would be both productive and effective – using what we call our Arcadia Impact Score. Our main focus is on three outcomes – total cost of care, avoidable ED visits and avoidable hospital admissions.
A physician’s clinical judgement is still key
No matter how good an algorithm may be at identifying actionable patients, no system can substitute entirely for clinical judgement given the complex nuances of any given patient. Therefore, I encourage ACO leaders to think of algorithmically generated lists as a starting point for consideration for care management. They should always be combined with the provider’s knowledge of the patient and their clinical judgement. I am confident that our current enhanced tools allow ACOs to accurately identify patients for effective care management. And we will keep improving our methods!