If you are a healthcare provider or payer organization contemplating an initial implementation of a Business Intelligence (BI) Analytics system, there are several areas to keep in mind as you plan your program. The following key components appear in every successful BI Analytics program. And the sooner you can bring focus and attention to these critical areas, the sooner you will improve your own chances for success.
Key Program Components
Last time we reviewed the primary, top-level technical building blocks. However, the technical components are not the starting point for these solutions. Technical form must follow business function. The technical components come to life only when the primary mission and drivers of the specific enterprise are well understood. And these must be further developed into a program for defining, designing, implementing and evangelizing the needs and capabilities of BI and related analytics tuned to the particular needs and readiness of the organization.
Key areas that require careful attention in every implementation include the following:
We have found that healthcare organizations (and solution vendors!) have contrasting opinions on how best to align the operational data store (ODS) and enterprise data warehouse (EDW) portions of their strategy with the needs of their key stakeholders and constituencies. The “supply-driven” approach encourages a broad-based uptake of virtually all data that originates from one or more authoritative source system, without any real pre-qualification of the usefulness of that information for a particular purpose. This is the hope-laden “build it and they will come” strategy. Conversely, the “demand-driven” approach encourages a particular focus on analytic objectives and scope, and uses this focus to concentrate the initial data uptake to satisfy a defined set of analytic subject areas and contexts. The challenge here is to not so narrowly focus the incoming data stream that it limits related exploratory analysis.
For example, a supply-driven initiative might choose to tap into an existing enterprise application integration (EAI) bus and siphon all published HL7 messages into the EDW or ODS data collection pipe. The proponents might reason that if these messages are being published on an enterprise bus, they should be generally useful; and if they are reasonably compliant with the HL7 RIM, their integration should be relatively straightforward. However, their usefulness for a particular analytic purpose would still need to be investigated separately.
Conversely, a demand-driven project might start with a required set of representative analytic question instances or archetypes, and drive the data sourcing effort backward toward the potentially diverging points of origin within the business operations. For example, a surgical analytics platform to discern patterns between or among surgical cost components, OR schedule adherence, outcomes variability, payer mix, or the impact of specific material choices would depend on specific data elements that might originate from potentially disparate locations and settings. The need here is to ensure that the data sets required to support the specific identified analyses are covered; but the collection strategy should not be so exclusive that it prevents exploration of unanticipated inquiries or analyses.
I’ll have a future blog topic on a methodology we have used successfully to progressively decompose, elaborate and refine stakeholder analytic needs into the data architecture needed to support them.
In many cases, a key objective for implementing healthcare analytics will be to bring focus to specific areas of enterprise operations: to drive improvements in quality, performance or outcomes; to drive down costs of service delivery; or to increase resource efficiency, productivity or throughput, while maintaining quality, cost and compliance. A common element in all of these is a focus on process. You must identify the specific processes (or workflows) that you wish to measure and monitor. Any given process, however simple or complex, will have a finite number of “pulse points,” any one of which will provide a natural locus for control or analysis to inform decision makers about the state of operations and progress toward measured objectives or targets. These loci become the raw data collection points, where the primary data elements and observations (and accompanying meta-data) are captured for downstream transformation and consumption.
For example, if a health system is trying to gain insight into opportunities for flexible scheduling of OR suites and surgical teams, the base level data collection must probe into the start and stop times for each segment in the “setup and teardown” of a surgical case, and all the resource types and instances needed to support those processes. Each individual process segment (i.e. OR ready/busy, patient in/out, anesthesia start/end, surgeon in/out, cut/close, PACU in/out, etc.) has distinct control loci the measurement of which comprises the foundational data on which such analyses must be built. You won’t gain visibility into optimization opportunities if you don’t measure the primary processes at sufficient granularity to facilitate inquiry and action.
Each pulse point reveals a critical success component in the overall operation. Management must decide how each process will be measured, and how the specific data to be captured will enable both visibility and action. Visibility that the specific critical process elements being performed are within tolerance and on target; or that they are deviating from a standard or plan and require corrective action. And the information must both enable and facilitate focused action that will bring performance and outcomes back into compliance with the desired or required standards or objectives.
A key aspect of metric design is defining the needed granularity and dimensionality. The former ensures the proper focus and resolution on the action needed. The latter facilitates traceability and exploration into the contexts in which performance and quality issues arise. If any measured areas under-perform, the granularity and dimensionality will provide a focus for appropriate corrective actions. If they achieve superior performance, they can be studied and characterized for possible designation as best practices.
For example, how does a surgical services line that does 2500 total knees penetrate this monolithic volume and differentiate these cases in a way that enables usable insights and focused action? The short answer is to characterize each instance to enable flexible-but-usable segmentation (and sub-segmentation); and when a segment of interest is identified (under-performing; over-performing; or some other pattern), the n-tuple of categorical attributes that was used to establish the segment becomes a roadmap defining the context and setting for the action: either corrective action (i.e. for deviation from standard) or reinforcing action (i.e. for characterizing best practices). So, dimensions of surgical team, facility, care setting, procedure, implant type and model, supplier, starting ordinal position, day of week, and many others can be part of your surgical analytics metrics design.
Each metric must ultimately be deconstructed into the specific raw data elements, observations and quantities (and units) that are needed to support the computation of the corresponding metric. This includes the definition, granularity and dimensionality of each data element; its point of origin in the operation and its position within the process to be measured; the required frequency for its capture and timeliness for its delivery; and the constraints on acceptable values or other quality standards to ensure that the data will reflect accurately the state of the operation or process, and will enable (and ideally facilitate) a focused response once its meaning is understood.
An interesting consideration is how to choose the source for a collected data element, when multiple legitimate sources exist (this issue spills over into data governance (see below); and what rules are needed to arbitrate such conflicts. Arbitration can be based on: whether each source is legitimately designated as authoritative; where each conflicting (or overlapping) data element (and its contents) resides in a life cycle that impacts its usability; what access controls or proprietary rights pertain to the specific instance of data consumption; and the purpose for or context in which the data element is obtained. Resolving these conflicts is not always as simple as designating a single authoritative source.
Controlling data quality at its source is essential. All downstream consumers and transformation operations are critically dependent on the quality of each data element at its point of origin or introduction into the data stream. Data cleansing becomes much more problematic if it occurs downstream of the authoritative source, during subsequent data transformation or data presentation operations. Doing so effectively allows data to “originate” at virtually any position in the data stream, making traceability and quality tracking more difficult, and increasing the burden of retaining the data that originates at the various points to the quality standard. On the other hand, downstream consumers may have little or no influence or authority to impose the data cleansing or capture constraints on those who actually collect the data.
Organizations are often unreceptive to the suggestion that their data may have quality issues. “The data’s good. It has to be; we run the business on it!” Although this might be true, when you remove data from its primary operating context, and attempt to use it for different purposes such as aggregation, segmentation, forecasting and integrated analytics, problems with data quality rise to the surface and become visible.
Elements of data quality include: accuracy; integrity; timeliness; timing and dynamics; clear semantics; rules for capture; transformation; and distribution. Your strategy must include establishing and then enforcing definitions, measures, policies and procedures to ensure that your data is meeting the necessary quality standards.
The data architecture must anticipate the structure and relationships of the primary data elements, including the required granularity, dimensionality, and alignment with other identifying or describing elements (e.g. master and reference data); and the nature and positioning of the transformation and consumption patterns within the various user bases.
For example, to analyze the range in variation of maintaining schedule integrity in our surgical services example, for each case we must capture micro-architectural elements such as the scheduled and actual start and end times for each critical participant and resource type (e.g. surgeon, anesthesiologist, patient, technician, facility, room, schedule block, equipment, supplies, medications, prior and following case, etc.), each of which becomes a dimension in the hierarchical analytic contexts that will reveal and help to characterize where under-performance or over-performance are occurring. The corresponding macro-architectural components will address requirements such as scalability, distinction between retrieval and occurrence latency, data volumes, data lineage, and data delivery.
By the way: none of this presumes a “daily batch” system. Your data architecture might need to anticipate and accommodate complex hybrid models for federating and staging incremental data sets to resolve unavoidable differences in arrival dynamics, granularity, dimensionality, key alignment, or perishability. I’ll have another blog on this topic, separately.
You should definitely anticipate that the incorporation and integration of additional subject areas and data sets will increase the value of the data; in many instances, far beyond that for which it was originally collected. As the awareness and use of this resource begins to grow, both the value and sensitivity attributed to these data will increase commensurately. The primary purpose of data governance is to ensure that the highest quality data assets obtained from all relevant sources are available to all consumers who need them, after all the necessary controls have been put in place.
Key components of an effective strategy are the recognition of data as an enterprise asset; the designation of authoritative sources; commitment to data quality standards and processes; recognition of data proceeding through a life cycle of origination, transformation and distribution, with varying degrees of ownership, stewardship and guardianship, on its way to various consumers for various purposes. Specific characteristics such as the level of aggregation; the degree of protection required (e.g. PHI); the need for de-identification and re-identification; the designation of “snapshots” and “versions” of data sets; and the constraints imposed by proprietary rights. These will all impact the policies and governance structures needed to ensure proper usage of this critical asset.
Are you positioned for success?
Successful implementation of BI analytics requires more than a careful selection of technology platforms, tools and applications. The selection of technical components will ideally follow the definition of the organizations needs for these capabilities. The program components outlined here are a good start on the journey to embedded analytics, proactively driving the desired improvement throughout your enterprise.