DaVita® Medical Insights


The health and safety of our patients and teammates is our top priority. We are keeping a close eye on this situation and reinforcing the extensive infection control practices already in place to protect them. Click here to find videos and additional resources.

Predictive Analytics Transforming Kidney Care

Predictive analytics are transforming health care. In particular, it helps us get the right treatment to patients at the right time. As the collection of therapies grows, it is important to target those therapies in efficient ways that will maximize the benefit to patients, payors and providers. Predictive analytics give us an increasing sophistication in how we deliver care and therapies to patients. This post summarizes the Podcast: Predictive Analytics published on DaVita Medical Insights in Feb. 2020.

Historical perspective

Predictive analytics are an extension and acceleration of a paradigm in use for many years. The conditions allowing for the current state of predictive analytics include the following:

  • An unprecedented amount of databy historic standardsis at our fingertips. As electronic health records (EHRs) have increased in size and scope, more information is available in databases in codified fields, where it can easily be found and indexed. Also, as systems are becoming increasingly interconnected, clinicians can access doctor’s notes, laboratory results and payor claims—a level of information that was not readily available in prior eras.
  • At the same time, we have an explosion of computing power. The computer can consider several thousands of data pieces. By looking at combinations and the timing of data events, the computer can identify patterns in powerful ways that allow for much more accurate and precise predictions, which when combined with clinical judgment are extremely effective.
Use in kidney disease

Kidney disease is one of the most fertile fields for the development of predictive analytics for a few reasons:

  1. Large datasets (on masses of people) are needed to be able to uncover unknown correlations and identify hidden patterns. Kidney disease is very common. Thirty million adults in the United States have some form of kidney disease, known and unknown, and that provides a lot of data to utilize.
  2. Kidney disease and associated outcomes from predictive analytics are relatively objective. A clinician can determine if someone has kidney disease by just looking at lab results. Teaching the models to look for the patterns is much easier when the outcomes of interest are easy to define and locate in the data. Finding a physician’s observation from a physical exam recorded in myriad ways and places in an EHR is much more difficult than it is to find a lab result or a hospitalization event.
  3. Health events, such as hospitalization and death, are common in patients with kidney disease. Events that are common make it easier to train predictive models to identify the patterns that can then pre-stage those events.
Opportunities for patients, payors and providers

Predictive analytics allows identification of which patient needs a particular therapy and when they need it by stratifying patients into different tiers. For instance, if a patient is at low-risk for developing end-stage kidney disease (ESKD), the patient may only require small changes to his/her diet, whereas a high-risk patient may require more intensive lifestyle changes. Depending on the level of risk for each patient (e.g., high-risk, moderate-risk and low risk), a specific care plan will be created and implemented. Thus, predictive analytics helps us determine the right treatment for individual patients at the right time.

From the standpoint of a payor, whether an insurer or a government program, they want to know several things to improve resource efficiency. First, they want to know that the resources allocated to a population are being used to maximize benefit in that population. Second, they want to know that they will get a return on their investment, meaning resources will not be outlaid if they don’t have a benefit downstream. Third, they want the combination of number one and number two to prevent future events that will drive up their outlay of costs. So, if you can prevent one hospitalization now, Medicare could potentially be saved tens of thousands of dollars two months from now. An investment now to help avoid a larger expenditure in the future seems worth it. From a payor standpoint, these make the delivery of health care and the use of their resources more efficient.

Providers, including doctors, nurse practitioners, physician assistants and nurses, live in a world sandwiched between patients and health insurers. Predictive analytics provide a way of getting everyone moving in the same direction. It simplifies the provider’s job in terms of needing to escalate the level of care for some patients and not others.

An ideal and successful model

The most important piece for a model is that it be accurate and reproducible. That is, the model will work on a population of Medicare Advantage patients who are mostly in their late 60s or 70s, equally well as in a health care insured group of patients who are in their 40s and their 50s. It will work as well in an urban environment as in a rural or suburban environment.

The second most important piece is that the model be built for a very detailed use case. The elements that need to be included are the outcome most desired to be known from the model and the timeline over which that outcome is likely to occur. For example, a model exists that predicts the progression risk of CKD to ESKD. Now, it is important to know if a patient is going to progress, but as a clinician, if that patient’s progressing to ESKD in 10 years that implies one set of interventions for that patient. However, if that patient is going to progress quickly over the next 10 months, a different suite of interventions is needed. In the second scenario, the patient needs to be educated about modalities, including counseling about transplant if appropriate; and they need dietary counseling. Specifying the time horizon over which the patient is likely to progress becomes critically important.

The third piece, which is important not to overlook, is that when you build a model, the data environment in which the model will be used needs to be understood. When DaVita builds its models, the first thing we do is work with our partners to understand what the data environment will be like for the actual use of the model. Then, we build in everything that users will have access to and nothing that they won’t. If something changes and there is greater access to more data, we can always go back and build a later iteration of the model. It is crucial to make sure that those models are practically implemented in the use cases so that they can have their intended consequences. For example, one may be able to feed into and build a model that expects lab data, EHR data, data on social determinants of health, and claims data to get a maximally accurate prediction. However, if in my use case, data on social determinants of health and claims data will not be available to the user in real-time, then that super accurate model will not actually work when needed.

Successful use of predictive analytics is the ability to minimize unwanted future events and maximize future health without unnecessary expenditure of resources. We want to be great population health managers, great clinicians and good stewards of health care resources.

DaVita built a predictive model for patients on dialysis that helps predicts their 90-day risk of hospitalization that in testing it in many separate datasets proved to be highly reproducible. However, it’s unknown if the model is successful without a suite of interventions linked to the model’s predictions.

By using output from the model along with clinicians’ judgment, patients predicted to be at high-risk for hospitalization received focused and individualized care that included additional interventions. These patients had hospitalization rates deflecting downward. The medium-risk patients who received incremental services beyond standard of care, but not as many interventions as the high-risk patients, also had their hospitalization rates deflect downward. Patients at average risk for the population had no increased rate of hospitalization. In turn, this unwanted future event, this hospitalization, was often prevented in the patients who were the most vulnerable, without adversely affecting the other patients.

Supporting value-based care arrangements

Value-based care is when the financial risk for patients is transferred in whole or in part from the traditional payor to the provider organization. It is important that these organizations, as they take on risk, maximize the likelihood of being efficient and effective with their resources. This allows them to hedge against the likelihood of achieving the downside of that risk and the likelihood of achieving the upside of that risk. It not only supports the provider organizations, but it also supports the payors, because that’s the factor that allows them to shift that risk to the providers. Most importantly, it supports the patients because by aligning the physicians, provider organizations and the payors all to the same true north, the collective team can help patients stay healthy and out of the hospital. In value-based care, predictive analytics are increasingly used as a powerful tool that allows all players to drive in the same direction toward shared goals.

The future of predictive analytics                                   

The future is almost limitless. It is plausible that with the efficiency and the effectiveness that predictive analytics engender, they will not be restricted to just the models that exist today. As predictive analytics are more integrated into clinical processes and as predictions are married to tangible steps to take for patients, the use of predictive models can be expanded to drug dosing, for example. Predictive analytics may allow the health care industry to better determine the right dose more quickly than previous methods and avoid ineffectiveness, adverse events or serious risks for the patient.

To end with a note of caution: predictive analytics are not a substitute for clinical judgment. Predictive analytics are a tool in the same way that a radiology test or a lab test are. Each of these “tools” supply information to a health care provider that has to be filtered through that provider’s judgment. Although predictive analytics are very powerful, ultimately, the care of patients rests with the expertise and the judgment of the provider. Moving forward, remember that these models have limitations to what they can do while having a huge upside to what they promise.

Steven M. Brunelli, MD, MSCE

Steven M. Brunelli, MD, MSCE

Steven M. Brunelli, MD, is vice president and medical director of health analytics and insights at DaVita Clinical Research. Before joining DaVita, he was a faculty member at Harvard Medical School and the Brigham and Women’s Hospital, where he directed an active clinical research group that focused on chronic kidney disease, pharmacoepidemiology and pharmacoeconomics, dialysis outcomes epidemiology and the hospital’s dialysis service. Dr. Brunelli also served on the American Society of Nephrology’s Dialysis Advisory Group and its Comparative Effectiveness Taskforce. He has published more than 120 peer-reviewed articles, and serves on editorial boards at the Journal of the American Society of Nephrology, the Journal of Nephrology and the American Journal of Kidney Diseases. He completed medical school and earned a master of science degree in clinical epidemiology at the University of Pennsylvania.