By Michael Simon, PhD
Care management programs are becoming an increasingly essential means for provider groups to maintain a high quality of care while also controlling costs and utilization under value-based agreements.
The problem many providers face in launching care management programs, however, is the heavy use of resources that they require. For example, each nurse care manager typically manages a panel of between 50 and 200 patients, while each enrolled patient generally spends between three and six months in a program.
For health care organizations (HCOs), the challenge lies in determining how to most efficiently allocate care managers’ limited time toward the needs of each individual patient. Generally, HCOs have employed algorithmic models to identify the patients who are the top candidates for care management programs based on future likelihood of high utilization. However, these algorithms are sometimes one-dimensional and fail to take into account whether patients are likely to respond well to care management. As a result, care management programs sometimes do not enroll the patients who would stand to benefit from them the most.
Now, with more fully developed machine-learning techniques and more comprehensive data sets, provider organizations can use predictive analytics to pinpoint the patients who are likely to respond best to care management programs.
More Than Just the ‘Sickest’ Patients
Hospitals and health systems are increasingly emphasizing care management programs for a simple reason—care management has been shown to improve outcomes and enhance patient care. With comorbidities and complex conditions, many of today’s patients are too challenging for one physician to handle alone. Experienced and knowledgeable care managers are capable of establishing a rapport with their patients to inspire behavior changes that lead to fewer admissions and emergency department visits, less utilization, and better outcomes.
The most common drawback of care management programs is that they often don’t serve the right patients. The traditional approach is generally limited to detecting the “sickest” patients—at the expense of other important factors such as whether the patient’s condition is likely to respond well to care management or whether the patient’s history suggests he is likely to change his behavior in accordance with care management.
The problem with this traditional approach is that it overleverages utilization data without accounting for the fact that all high-utilizers are not “actionable” patients—those who have a high likelihood of benefitting from a care management program. For example, a patient with end-stage renal disease on dialysis may be very ill with a high level of utilization but would realize little positive impact from interventions through a care management program. This approach to care management patient enrollment results in a wide range of patient types, which may lead to a misallocation of scarce care management resources.
Unlike the traditional approach to determining whether a patient is a strong candidate for care management, predictive analytics leverages a much wider array of clinical and demographic data, such as condition typologies and social determinants of health. With the help of integrated data sets that expand beyond typical data resources, predictive analytics can help health systems easily sort through multitudes of patient records to separate actionable patients from those unlikely to respond to care management.
Creating New Models for Care Management
One of the most important components of using analytics to improve care management involves “training” the predictive model using a large and diverse data set that includes outcomes that align with an HCO’s specific goals. For example, if the objective is to reduce avoidable readmissions and emergency department visits, the data should include a wide range of patients with all available outcomes to help the model identify patterns and trends that predict potential future utilization.
Next, a machine learning algorithm iterates quickly through increasingly precise models, which can take into account a diverse range of data, including morbidity risk, frailty concepts, environmental factors (such as local housing vacancy and high school graduation rates), clinical observations (such as blood pressure and body mass index), and health care utilization statistics. The objective is to identify a model that best describes the relationship between these diverse inputs and the outcome variables. Developing these types of predictive models can require a significant investment of time and resources, which is why analytics tools that align with HCO requirements are a useful resource in a care management strategy.
Due to their ability to accurately project utilization, cost, and outcomes, predictive analytics solutions can identify the most “impactable” patients—that is, those for whom care management would be most productive and effective. Care management teams can then use this information to design programs that optimally allocate their organizations’ resources.
A Complement to Clinician’s Judgment
Predictive analytics delivers a robust and efficient means of identifying patients to target for care management, but what it doesn’t provide is a replacement for the medical judgment, not to mention intuition, of an experienced clinician. Nonetheless, by helping to spot patients who may become high-cost utilizers before clinicians are able to, predictive analytics solutions represent a complement to a clinician’s judgment that can help HCOs improve population health.
— Michael Simon, PhD, is a data scientist at Arcadia.