Reducing Readmissions w/ Business Intelligence & Analytics – Part 1
Re-hospitalizing a patient is costly, potentially harmful to the patient, and often avoidable. And for certain hospitals, a patient readmitted within 30-days of discharge results in significant financial penalties. Given patient safety and financial impacts, healthcare organizations are using analytics as a means of reducing readmissions after initial discharge.
This post presents insights and lessons learned from Andrew Satz titled “Data Analytics to Predict When a Patient Will Readmit”. In future parts of this blog series, we’ll share insights and ideas on using analytics and technology to reduce readmissions from other subject matter experts:
- Part 2: Hospital Readmissions – Michele Russell of Russell Consulting Group
- Part 3: Dashboard to Forecast Healthcare Outcomes – Kevin Oppenheimer of KGO Consulting
- Part 4: A Care Transition System to Reduce Hospital Readmissions – Ed Kirchmier of AAJ Technologies
Data Analytics to Predict When a Patient Will Readmit
Andrew Satz Shares Insights on Reducing Readmissions
We use data science and artificial intelligence to improve patient outcomes, heighten quality of care and prevent financial losses. And today I want to talk a little bit about correlation and causation, the difference between them and how they relate to readmissions. In order to really understand the difference between correlation and causation, I’m going to give you a short example.
Correlation is Not Causation
There was a study done that clearly demonstrated that people who have [cigarette] lighters in their pockets had a higher chance of lung disease. And while we know a lighter in a pocket might be associated or correlated with lung disease, we know, in fact, that lighters in a pocket are not the cause of lung disease. In fact, one major cause of lung disease is smoking cigarettes. So, while lighters are correlated with lung disease, they are not necessarily the cause of lung disease. And how we determine the difference between causation and correlation relies heavily on intuition; but it also needs to be supported by mathematics and statistics.
A similar but surprising example of correlation is with hospitalization from pneumonia. But most of those readmissions are not because of pneumonia related causes. So, when you look at your data, you’ll see a big impact on your readmission correlated to the original admission of pneumonia. And that’s a correlation. And while the pneumonia is correlated with the readmissions, it isn’t necessarily the cause. In fact, this study which came out of the National Institutes of Health show that it’s the comorbidities that are causal factors that lead to the readmission.
So, in reality, while pneumonia is highly correlated, it’s not the cause of the readmission. It’s important that you look deeply at data; whether by examining the relationships between patients within a certain original admission or other factors about patients that could lead to readmissions that have nothing to do with the original cause of admission. Or understand complementary causes of readmissions like comorbidities. And that’s where analytics, data science and machine learning can be useful.
Identifying Key Data Features Impacting Business Intelligence
Some of the data you can look at when you’re trying to understand your patient is the data in your patient records such as what the patient was admitted for, and the doctor attending the patient. But their records also include things like age and weight, where they live, and some of the conditions that they have. When you start looking at these, we get some really deep dimensions about the patient and it becomes highly complicated to look very deeply about patients. And clustering them so that you can understand a relatable cause. In fact, adding just one additional feature like sex or weight or a single comorbidity makes understanding the cause of a readmission even more complicated.
Reducing Readmissions by Limiting Features
We have to find ways to make sure the data doesn’t go from this small data set about eight patients into something that blows up to a point where you have a ton of useless information. And we can use machine learning to determine what features about these patients are important. Let me be clear about the word feature: a feature is just another piece of input that we can use to describe or predict.
For example, a feature about a patient could be their age or their weight. When we’re looking at large datasets, we want to start to eliminate features that provide little or no information when making a prediction. From the later lung disease example, I gave earlier, we can use math and statistics to eliminate features that have nothing to do with lung disease; like shoe size. Computers and mathematics can be used to process large amount of data to determine how important a feature is in predicting readmissions. We can use analytics and data science to understand what features are important when it comes to cause or correlated features of readmissions. And now we can use this information to start classifying patients and predicting which patient will readmit; as well as a list of features that are associated with readmissions.
What this means is that you can use the features associated with historical readmissions to deliver medical health or social interventions that can predict readmissions.
Real World Example: Using AI for Reducing Readmissions
Now let’s go over a real-world example of how data analytics is being used today to understand correlated and causal features in order to predict and understand readmissions. The healthcare system from this example somehow overcame the incredibly challenging task of combining their data including clinical data, their financial data, administrative data, patient experience data and more. They used what I just mentioned – sort of feature selection – to figure out what was both correlated and causally related to the readmission.
These points included items like:
- Whether the patient was a smoker
- Where the patient lived
- The financial risk involved with a particular procedure
- If the patient had a primary care physician
- How old the patient was
- How the patient was admitted
- The medications the patient was on
- And a bunch of other data
With that clean, organized and properly stored data, a data scientist built machine-learning algorithms to model readmissions in order to predict whether a patient would return to the health care system within 30 days; as well as the features that had the most correlation and causation with readmissions.
If you actually want to know more about some of these machine learning models and how they can be used, you should go to AAJ’s YouTube page and watch the video I did on artificial intelligence in the Future of Healthcare Forum organized by AAJ Technologies.
The Results of Using Machine Learning to Reduce Readmissions
The resulting predictions and understanding from these machine learning algorithms were really surprising, especially to those who were experienced healthcare professionals. What they found was that patients who took more medications were actually at less of a risk of readmissions than those who took no medications. When they dug deeper they found that people on no meds had several untreated conditions, which increases the likelihood that a patient would be readmitted. There was a similar pattern when it came to age. There was an assumption that older patients are going to have a higher likelihood of readmitting. But in fact, it was the younger patients who were more likely to readmit because they didn’t have a primary care physician; or they weren’t tracking themselves as regularly to find out if they had healthcare problems.
That’s how we actually use artificial intelligence to reduce readmissions and some of the machine learning algorithms and the feature engineering that we can use to actually predict whether readmission will happen.
More Insight and Actionable Information on Reducing Readmissions
This content is originally from our “Reducing Readmissions with Business Intelligence & Analytics” webinar held earlier this year. You can view the PowerPoint presentation here.
You can also reach out to Murray Izenwasser, our Vice President of Digital Transformation, for more on how AAJ Technologies can help your company leverage BI and analytics for reducing readmissions by using the form below.
The post Reducing Readmissions w/Business Intelligence & Analytics – Part 1 appeared first on AAJ Technologies.