At Sanford Health, we have embraced the importance of data and analytics in improving the health of our communities. In order to accomplish this mission, we needed to first establish a strong foundation through centralizing our analytics teams, creating a solid base of data governance, and instituting the “One Source of Truth” through a forward-looking data warehouse strategy. We have learned numerous lessons from making this transformation: perhaps the most unexpected one was the importance of prescriptive, rather than predictive, analytics.
Analytics for the Value-Based World
Sanford Health is an integrated health care delivery system based in the upper Midwest serving 2.74 million people in 300 predominately rural communities across 250,000 square miles. The system is composed of 45 medical centers and 289 clinics served by more than 1,400 employed physicians. Annual revenue is in excess of $4 billion including more than 180,000 health plan members as part of the system.
The transition from a fee-for-service world into a value-based world clarified for us the importance of data and analytics. We had to deepen our understanding of the costs and quality of our care, and how that care fit into the larger context of our patients’ lives. While fee-for-service models still predominate in our area, we are pushing our payers to adopt value-based payment models because we believe value-based care will benefit our patients, and creating analytics that prove its value is a key part of our mission.
In the summer of 2015, Sanford Health established the new function of Enterprise Data and Analytics (EDA), whose mission is to generate value and improve care through strategic application of data and analytics. EDA merged two legacy groups: Decision Support and the report developers from the IT infrastructure. Historically, these two groups had significant overlap in function but limited coordination. Both pulled data from disparate data systems and created analytics outputs for different purposes, but problems often arose because these approaches yielded different and sometimes contradictory answers. Additional legacy analytics teams were subsequently folded into EDA, including operational, health plan, and human resource analysts.
Upon the creation of a centralized analytics team, we were able to stratify into internal functions, including a technical team focused on data architecture and database design, and a small core of data scientists working on bleeding-edge technologies like cognitive computing (self-learning systems that mimic the way the human brain works). We also shifted our organizational chart so that the newly unified EDA reported to the corporate Chief Financial Officer rather than the Chief Information Officer (who was one level lower in the reporting structure and reported to the Chief Operating Officer). Data and Analytics became an active partner in setting and achieving overall organizational goals, rather than solely an operational support function.
Building the Foundation for “One Source of Truth”
Like most health care organizations, Sanford Health did not have “One Source of Truth”: a single data repository that all facets of the organization both contribute to and draw from. Such a repository requires rigorous data governance.
Our history was filled with mergers, and each one brought legacy processes and definitions that lingered. Consequently, even seemingly straightforward metrics such as length of stay were defined in multiple ways across the organization. Additionally, our data was heavily siloed. For instance, clinical quality data did not interface with patient satisfaction data, which did not interface with human resource data. As a result, it was impossible to analyze these data sets to determine how staffing and quality of care affected patient satisfaction. We needed to focus on creating a solid foundation for EDA before we could embark on targeted analytical projects.
This foundation included:
A centralized team: Centralizing the team helped standardize not just the analytic approaches, but also the approach to extracting data from Sanford’s many information systems, a process vital to setting a framework for analysis. The overall leadership for the EDA team, which comprises more than 60 members, is shared by two people: a seasoned business leader steeped in decision support who handles day-to-day internal EDA management, and a practicing physician with expertise in health services research, business, finance, and frontline clinical delivery who works with clinicians and top management to set the overall analytics strategy.
A data governance structure: A common language was vital for standardization of analytics and the effective application of analytic outputs. Without this language, organizational leaders can be rapidly paralyzed with inconsistent views of common metrics. The governance structure included both standard definitions for all data elements used and an organization-wide method of benchmarking our performance against external standards.
Within 12 months of starting, our data governance committee defined more than 150 discrete terms. It also adopted two commercial databases for standardized benchmarking in clinical, financial, quality, and operational metrics. (Previously, some clinical departments adopted specialty-specific sources of benchmarking data, creating inconsistencies that made organization-wide comparisons difficult or impossible. The commercial databases chosen enabled us to extract the same insights previously obtained through the specialty registries in a consistent way.)
A data warehouse strategy: We established a forward-looking data warehouse strategy. Historically, data resided in “silos” associated with each individual system. A standard data warehouse took the data out of the silos and transferred it to one unified data structure through a process called “extract-transform-load,” or ETL, which was time-consuming and needed to be redone every time a new data source was introduced or if data governance changed significantly. Instead, we created a “virtual” data warehouse, where the data remains in its original system and we draw from it as necessary. Because the silos are now covered by a single data governance structure, we can combine data sources more readily. This approach also allows us to rapidly incorporate new sources, such as patient-generated health data and socioeconomic data.
Leapfrogging Predictive Analytics for Prescriptive Analytics
Predictive analytics — the ability to use past data to anticipate and prepare for future events — is often the goal of a revamped and centralized analytics function. However, we discovered that predictive analytics is only a starting point for clinical and operational algorithm development. It’s not enough to predict an event; we must also use analytics to help us prescribe the appropriate response.
In our world of hospital medicine, for example, we know that readmission rates are an important metric for quality of care. Algorithms work best for segmenting risk and alerting the clinician, commonly in a numeric scale indicating risk, but they rarely help the clinician decide the best action to take. Do I alter the medications? If so, which ones? Do I keep the patient longer in the hospital? Should I change the discharge location from home to a rehabilitation facility? Identifying which one of these actions would have the greatest impact is a job for prescriptive analytics, a relatively immature science.
Clinicians can’t change some of the common factors that may predict readmission: the patient’s age, gender, diagnosis, and length of stay. But others — say, a missed clinic appointment, or a certain level of weight gain 3 weeks before a readmission — are potentially actionable. For these types of factors, we have two goals. First, we must determine which actions make a difference, and whether the difference is enough to change how our clinicians practice. (For example, if we make sure all patients attend all their clinic appointments, will there really be fewer readmissions?) Second, once analytics help determine those actions, we must communicate these data-driven insights to clinicians and help them incorporate the resulting changes in a way that improves care while minimizing disruption to the current patterns of practice.
Health care analytics is still a developing field, but we feel that we have built a strong foundation and are on the right path. We look forward to sharpening our ability to predict patterns and events based on our “One Source of Truth,” and to prescribe the most appropriate interventions as we improve the care and health of our communities.