What Is Data Mining in Healthcare?
a risk model (based on comorbidity, severity score, physician scoring, and other factors) to patients in the census, run the data through regression analysis, and assign a risk score to each patient. The health system uses this score to inform which care-path patients take after discharge so that they receive the appropriate follow-up care.
Although these predictive models require a committed cross-functional team (physicians, technologists, etc.) and need to be tested over time, these clients are happy with the progress and preliminary results. They are moving beyond the theory of data mining into real, pragmatic application of this strategy.
Using Analytics to Track Fee-for-service and Value-based Payer Contracts
Let’s go into more depth about how one of these clients is using data mining and predictive analytics to address a major trend in healthcare today: effecting a smooth transition from fee-for-service (FFS) to a value-based reimbursement model.
We all know that the transition to value-based purchasing is happening. It represents the future of healthcare. But this shift isn’t a switch that can be flipped overnight. Instead, health systems must juggle both care delivery models simultaneously and will likely have to do so for many years to come.
We are working with a team at a large, nationally recognized integrated delivery network (IDN) that is using data mining to help navigate this transition—working to succeed in risk-based contracts while still performing well under the fee-for-service reimbursement model. This means that they need to lower their census for patients under risk contracts, while at the same time keeping patient volume steady for patients not included in these contracts.
Monitoring and Predicting Fee-for-service Volumes
A significant percentage of this IDN’s revenue comes from out-of-state referrals to its top-rated facilities. The team wants to ensure that these FFS contracts remain in place and supply a steady stream of business. To monitor this process, they have implemented an enterprise data warehouse (EDW) and advanced analytics applications. The EDW aggregates multiple data sets—payer, financial, and cost data—and then displays dashboards of information such as case mix index (CMI), referral patterns for each payer, volumes per payer, and the margins associated with those payers.
This system enables the team to mine data viewing trends in volume and margin from each payer. At this point in the implementation, the team is able to see within a quarter—rather than after a year or two—that referrals from a certain source are slowing down. They can then react quickly through outreach, advertising, and other methods.
As you can see, this innovative system we’re developing is still one that is reactive—though it identifies trends quickly enough that the health system can react before the margin takes much of a hit. But we are currently refining the system to become one that is truly predictive: one that uses sophisticated algorithms to forecast decreases in volume or margin by each referral source.
Participating in Shared-risk Contracts
Of course, at the same time as they work to optimize referral volumes, the health system’s team must also manage patients in value-based contracts. The IDN is an accountable care organization (ACO) with shared-risk contracts that cover tens of thousands of patients. Just as they are bringing referrals into the hospital, they are optimizing care to keep their at-risk population out of the hospital. They are, therefore, also using the EDW to help ensure that patients in this population are being treated in the most appropriate, lowest-cost setting. Analytics enables the team to monitor whether care is being delivered in the appropriate setting, identify at-risk patients within the population, and ensure that those patients are assigned a care manager.
Health systems nationwide are feeling the pressure of figuring out how to straddle the FFS and value-based worlds until the flip is switched. Having the data and tools on hand to predict their volumes and margins—while managing value-based contracts using the same analytics platform—is giving a significant advantage.
Pragmatic Application of Data Mining to Population Health Management
Another client is using the flexibility of its EDW to concurrently pursue multiple population health management initiatives on a single analytics platform. We are working together on two initiatives that employ the EDW, advanced analytics applications, and data mining to drive better management of the populations in the health system’s clinics.
Data Mining to Improve Primary Care Reporting
The first initiative mines historical EDW data to enable primary care providers (PCPs) to meet population health regulatory measures. This clinic’s PCPs must demonstrate to regulatory bodies that they are giving the appropriate screenings and treatment to certain populations of patients. Their focus to date has been on A1c screenings, mammograms for women over 40, and flu shots. The EDW and analytics applications have enabled the PCPs to track their compliance rate and to take measures to ensure patients receive needed screenings.
The Health Catalyst Advanced Application for Primary Care shows trending of compliance rates and specific measurements over time. So, the clinic can view how a patient’s A1c or LDL results are trending. They also see patients who may still be in a healthy range but over the last 18 months are trending closer and closer to an unhealthy result, then proactively address the issue.
A fun story from this clinic involves a Nurse Practitioner who joined the practice 20 years ago with a dream of changing the standard of care for diabetes. She tried to create concise reports but ran into one roadblock after another and finally resorted to spreadsheets mapped to EMR fields as a reporting mechanism, realizing it’s a less-than-ideal stopgap. Finally, after 20 years, her dream came true with the Health Catalyst solution to deliver monthly reports to individual physicians showing their diabetic patients and respective compliance to the standard of care.
Having this data readily on hand has also enabled the clinic to streamline its patient care process—enabling front-desk staff and nurses to handle screening processes early in a patient visit (which gives the physician more time to focus on acute concerns during the visit). This approach allows physicians to see more patients and devote more time to those patients’ immediate concerns. And it allows each member of staff to operate at the top of his or her license and training.
Data Mining to Predict Patient Population Risk
The second initiative involves applying predictive algorithms to EDW data to predict risk within certain populations. This process of stratifying patients into high-, medium- or low-risk groups is key to the success of any population health management initiative. Interestingly, some patients carry so much risk that it would be cheaper to pre-emptively send a physician out to make a house call rather than waiting for that patient to come in for a crisis appointment or emergency room visit. The clinic needed to be able to identify these high-risk patients ahead of time and focus the appropriate resources on their care.
To better risk stratify the patient populations, we applied a sophisticated predictive algorithm to the data. Using the data, we identified the clinical and demographic parameters most likely to predict a care event for that specific population. We then ran a regression on the clinic’s historical data to determine the weight that should be given to each parameter in the predictive model.
By applying such a tailored algorithm to the data, the clinic has been able to pinpoint which patients need the most attention well ahead of the crisis. Importantly, the clinic has integrated this insight into its workflow with a simple ranking of priority patients. This allowed for development of improved processes for managing the care of at-risk patients. For example, each week the physicians and care coordinators discuss the risk level of each patient with an appointment scheduled for that week. They can then create a care management plan in advance to share with the patient during the visit.
The clinic also looks at Patient Activation Measure® (PAM) scores and uses that data to determine patient engagement and activation. This leads to shared decision-making between the PCP and the patient, as the physician is able to determine ahead of time those patients who are at higher risk for non-compliance or might be unable to fully participate in their care.
Data Mining to Prevent Hospital Readmissions
Reducing 30- and 90-day readmissions rates is another important issue health systems are tackling today. We have used data mining to create algorithms that identity those patients at risk for readmission.
When your health system has an adequate historical data set—i.e., you have adequate data about