Why: Inferential Analytics

We previously defined predictive analytics as a branch of analytics providing insight into the future. We have a defined outcome, such as a birth or the onset of a hospital acquired illness such as sepsis, and predict the occurrence using statistical and machine learning techniques. However, this doesn’t allow for the study of why the outcome is expected. It is often this understanding of why an outcome is expected that makes clear the appropriate clinical or operational action. For this, we need to employ a set of tools and analytical methods called inferential analytics.

Understanding Why a Patient Is at Risk of Harm

Suppose a panel of patients in an inpatient facility are at risk of some harm, such as a hospital acquired infection (HAI). It is often not enough to output a “yes/no” decision, or a probability that the patient will acquire said harm. Predicted outcomes can frequently contradict the clinical judgment and intuition of doctors who may have practiced medicine for decades. This leads to natural skepticism of model results: a challenge that must often be overcome early in the adoption of data science solutions in hospital environments. Understanding why a patient is at risk bolsters the credibility of the model significantly, often uncovering information that a clinician can easily overlook. For instance, one intuitively knows that a respiratory rate trending upwards over four hours is bad; however, temporal information like this can easily go unnoticed during hectic periods such as handoffs or transitions of care. As a result, the most meaningful information is used in patient care, and outcomes improve.

Inferring the Causes of a Costly Event

Prediction, while not always the end goal, can provide a convenient framework for root cause analysis. Previously, we discussed the notion of using process improvement techniques to identify potential causes for waste in an inpatient pharmacy. Here, the goal is not to predict whether a future dose will be wasted; rather, the goal is to classify wasted dispense events with their most likely cause. Once this is done, the predominant causes of waste can be reported and monitored. However, it stands to reason that a model with strong predictive power has strong inferential capabilities as well. When used for prediction, there is a natural ranking of features from most to least important, such that the model can prioritize certain features over others. This ranking has enormous inferential power; the feature with highest rank can reasonably be assumed the root cause. Ergo, inferential analytics forms the backbone of root cause analysis – a critical step towards pharmacy cost reduction.

Prediction vs. Inference Mechanics

The modeling technique used is of critical importance. For example, statistical models tend to model outcomes through mathematical equations. These allow end users to draw very simple mathematical inferences such as, “Holding all else fixed, surgical patients are 2.2 times more likely to experience a certain HAI compared to medical patients”. Machine learning methods generally model a response through algorithms. Many of them hold tremendous predictive power at the expense of end-user interpretability. However, certain algorithms such as Random Forests or Stochastic Gradient Boosting enable the user to see which features of the dataset are of greatest importance. Unfortunately, there is often a tradeoff between predictive power and inferential capability though this exchange can be tremendously fruitful.

It cannot be stressed enough that any data science solution requires an enormous amount of clinical buy-in and integration with workflow to be successful. With that said, understanding why an event happened in the past is paramount to preventing it in the future. When this principle is properly applied in the context of a well-designed data science solution, the impact on patient care and cost reduction is very powerful.


About the Author

Richard Bryant is a data scientist with experience using data analysis, predictive modeling, and discrete event simulations to drive innovative technology solutions.  Utilizing an M.S. in Applied Statistics from the University of Michigan and a Lean Six Sigma Green Belt certification, Richard has provided analytical support to healthcare process improvement projects in various hospital departments, such as radiology, obstetrics, and pharmacy.  Prior to his role at Improvement Path Systems, he provided statistical consulting services to the Michigan Surgical Quality Collaborative and to the educational technology provider Echo360.