Using Predictive Analytics to Improve Healthcare Outcomes. Группа авторов
traditional graphic."/>
Figure 1.2 Explained variance of CLABSI, traditional graphic.
Figure 1.3 Explained variance of CLABSI, bubble graphic.
Both models reveal that the location of the central line insertion predicted 2.2% of CLABSI, the assigned RN predicted 1.1%, and the phase of the project predicted .4%. It has been the experience of this author that the visual representation in Figure 1.2 does not convey information as quickly as does the visual representation in Figure 1.3, in part because it relies too heavily on statistical symbols and equations to convey the information. In this book, however, you will see that most of the graphic representations of models and their results are expressed in traditional graphs, as many readers are likely to want more information than bubble graphs convey.
Analysts can provide a review of the results, but it is only the staff members who can provide validation or reflections that may suggest the need for secondary analysis to understand the data more deeply. When the data is being presented, pay attention to the listeners' responses. Even people who do not want to speak up may provide useful insight through nonverbal responses such as silence or even a shift in energy in the room. All of these cues can be informative. It is not uncommon for this author to pull listeners aside to discuss the nonverbal cues or silence that was observed. When encouraged to express themselves, these are often people from whom extremely valuable feedback is elicited.
Step 13: Respecify (Correct, Refine, and/or Expand) the Measurement Model
This is also the work of the data analyst, but it is done in close collaboration with staff members. This step includes refinement and possible expansion of the model to make it an even more sensitive model to detect predictors accurately. During presentation of the data to staff members, new variables will be identified, or variables that could not be measured in the first round but belong in the overall model will be addressed. Respecify the model to include anything that could not be included in the initial analysis or was identified in the interpretation as missing, and delete anything that was determined in the analysis to be unimportant. In one real‐life example, we were measuring the performance of charge nurses as the variable of interest, and we had proposed that our three predictor variables were (a) demographics of the charge nurse, (b) attending the charge nurse program, and (c) the preceptor who trained the charge nurse into the role. The originally specified model looked like Figure 1.4.
After Model 1 was used to examine this variable of interest, and after the data was presented to unit managers, charge nurses, and staff members from the unit, those who attended the presentation reported that Model 1 was missing two influential predictor variables: (a) mentoring of the unit manager and (b) resources available on the job for charge nurses to execute their required role. These influential variables were added to a respecified model (Figure 1.5), and the study was conducted again to see whether analysis of the respecified model could further explain what was influencing performance of charge nurses.
Figure 1.4 Model 1 to measure new charge nurse performance.
Figure 1.5 Model 2, respecified with new predictor variables to measure new charge nurse performance.
Note also that in structural models such as Figures 1.4 and 1.5, we have rectangles that look like they are representing one variable, when in many cases they represent multiple variables. For example, the rectangle labeled “Demographics” in both figures might be representing a dozen or so variables. These smaller, more compact models, which appear throughout this book, are called over‐aggregated structural models. Remember when you see them that what looks like a model testing three or four variables is actually testing dozens of variables at the same time.
Step 14: Repeat Steps 2–13 if Explained Variance Declines
In nearly every instance, the data analyst will, along with staff, be repeating Steps 2–13. When initiating use of predictive analytics, conventional wisdom says that at least 50% of the variance should be explained using regression analysis, but it is the experience of this author that explained variance of 70–75% for a variable of interest can be achieved with a good fitting model, using 10 predictor variables or fewer, in a regression analysis.
As practice changes are implemented based on the information that emerges, variables from the initial model will no longer predict the variable of interest because the problem (or part of the problem) will have been solved by the practice changes. Traditionally, the analyst would then have to start over and develop a new model, but in this case, much of the work has already been done when developing the initial full model that is graphically depicted in Figure 1.1. As you return to Step 2, you will review the existing full model and rerun all the analytics to identify existing predictor variables that have now become an issue due to the new practice changes and/or identify new variables that relate to the variable of interest.
Step 15: Interface and Automate
Collecting data from a variety of software can take a lot of time, and it costs a lot of money for staff members to collect the data. These are compelling reasons to examine how technology can be used to automate the specified models developed to study the variable of interest. To examine outcomes in as close to real time as possible, interface the software and applications to one repository of data so the data can be examined as it comes in. Programs can be written to make the mathematical formula run every time one of the new variables comes into the dataset. A program can even be written for automatic respecification of the model as operations of clinical care improve. Manual respecification of a measurement model takes a lot of time, but if the program is set up to detect a fall in the explained variance for any variable, a program can be written that automatically reruns the correlations of all the variables in the model and then automatically builds a new model. Coefficients can be used to identify some specific aspects of how the newly added predictor variable is affecting the outcome variable. For example, if the variable of interest was CLABSI incidence, and the predictor variable is “central line type is causing infection,” the coefficients can identify what type of central line is causing infections, what unit/department it is most likely to occur in, and/or other specifics from other predictor variables in the model.
A program can be written for automatic respecification of the model as operations of clinical care improve.
Step 16: Write Predictive Mathematical Formulas to Proactively Manage the Variable of Interest
Over time, the analyses from models used to study how specific variables