The first step and arguably greatest challenge in healthcare data analysis is transforming incalculable quantities of messy raw data into an asset capable of producing measures that are timely, accurate, and clinically relevant. Successful completion of this first step will produce immediate insights from even simple descriptive statistics, such as counts, proportions, and averages.
At some point, however, researchers will want to move beyond these initial analyses to discover relationships between health measures or to address specific use cases, with a focus on addressing questions that are both high value and tractable. Such goals require more complex methods.
Ursa Studio gives users the ability to quickly perform complex analyses that are driven by their unique questions and business needs. Here, we describe some of the tools available for advanced analytics in Ursa Studio.
One way data can drive improvement is by showing decision-makers where to act. A basic analysis can demonstrate the levels and changes in performance outcomes. It can also show levels and changes in the variables believed to drive these outcomes.
However, a basic analysis does not provide explicit, quantitative information on the relationship between hypothesized drivers and outcomes. Instead, researchers typically need to rely on their intuition about these relationships when prioritizing variables for potential interventions.
Ursa Studio’s Driver Diagram allows researchers to test and hone their intuition about what drives outcomes by modeling the relationships in the data. Users can specify which upstream measures they believe will drive changes to downstream measures. They can also identify control variables by assigning them as Risk Stratification Fields in Measure Workshop.
When they run a report, a model is fit to the data with upstream measures serving as predictors and downstream measures as outcomes. The results allow users to test for the existence and strength of the associations between predictors and outcomes.
At the top of the Driver Diagram is a slider that allows users to explore hypothetical scenarios by manipulating the value of input measures to assess their impact on output measures. For example, one may find the predicted change in per member per month (PMPM) costs associated with increased enrollment in a population health management program designed to reduce costs while improving care.
And under the stronger assumption that the model is measuring causal effects, these hypothetical scenarios will represent the predicted outcomes from interventions.
Performance often varies markedly across different healthcare delivery units and programs, such as physicians, facilities, or intervention programs. One approach to performance improvement is to redirect patients from lower to higher performing units. One complication, however, is that this approach requires apples-to-apples comparisons, which are difficult to achieve. For example, Clinic A may have higher charges than Clinic B simply because it serves a higher risk population; if this is the case, transferring patients from Clinic A to Clinic B will produce no savings while potentially causing harmful disruptions for patients and providers.
Finding the variation across units, controlling for differences in case mix, and then predicting the potential impact of performance improvement are all necessary analytic steps. Opportunity Discovery allows users to quickly run this multi-step analysis across several measures.
To run Opportunity Discovery, a user makes three selections within Measure Workshop for:
When a user runs a report, Opportunity Discovery finds the best and the median performing units for each of the opportunity fields. Opportunity is then calculated as the improvements that would occur if the lower performing units were brought up to the level of the best performer.
For example, if the opportunity field is “health clinic,” then the opportunity is the cost savings, adjusting for case mix, that would accrue if all clinics reduced costs to the level of the best clinic. Users can also perform the same calculation for bringing all below-median clinics to the level of the median performer.
Both the Driver Diagram and Opportunity Discovery assume that the user knows which variables are relevant for an analysis. Sometimes, however, users might find themselves working in a new domain with many potential predictors. In these situations, a manual search through all variables can quickly become impracticable.
Ursa Studio’s Insight Discovery automates search over large data sets and multiple measures. The tool swiftly searches through all categorical predictors, as well as their combinations, to find those that explain the greatest variability in an outcome measure.
For example, if a data table contained the categorical predictors of sex and CKD stage, then Insight Discovery would test each of them as single predictors well as the combined predictor of “sex/CKD stage,” which would contain the categories of “female/CKD stage 1,” “male/CKD stage 1,” and so on. Original variables are referred to as “parent” predictors while their combinations are “child” predictors.
Insight Discovery is run from within Measure Workshop. The tool provides several options to customize the analysis:
Insight Discovery displays the best predictors in a series of charts that show how the outcome varies over the different category levels. Users may then use these findings to develop hypotheses about what drives performance and where there are opportunities for improvement.
The Driver Diagram, Opportunity Discovery, and Insight Discovery provide ready-made analyses to address some of healthcare’s most common questions. As organizations evolve in their knowledge, their questions naturally do as well, becoming more specific. Therefore, analyses must be customized to meet these needs.
Ursa Studio’s Advanced Analytics Models provide immense flexibility to choose the methods that will best achieve project goals. This feature includes linear and logistic regression models that are built in to Ursa Studio. Users can also upload bespoke Python or R code to implement more complex models. Since Python and R are the leading programming languages for machine learning and statistics, virtually any method is available as an Advanced Analytics Model.
Ursa clients have used this functionality to achieve a range of objectives, such as risk prediction, health provider rankings, natural language processing, and time-series forecasting.
To apply this feature, users navigate to the Advanced Analytics zone of Ursa Studio, where they can then attach a model to an Ursa measure or object. If they select a linear or logistic regression model, they will be prompted to input a set of predictors and their corresponding regression coefficients, as well as choose how to handle missing values; if it is a logistic model, they have the option to change the classification decision threshold. Applying a bespoke model requires a block of Python or R code that the user has written to perform a specific task.
Once a report is instantiated, the Advanced Analytics Model executes and outputs a table of results that may be incorporated into Ursa Studio’s standard reporting tools. Like other Ursa objects, Advanced Analytics Models maintain a full revision history for audit purposes and allow for one click recovery of previous versions.
To learn more about Ursa Studio’s Advanced Analytics suite, reach out to as at firstname.lastname@example.org.