The US Department of Health and Human Services announced in January 2021 that Home Health Value-Based Purchasing was set to be expanded from the initial nine pilot states to a national audience. The final rules for the program aren’t set to be defined until January 2022.
Even still, with this announcement, home health agencies are feeling the pressure to find innovative and unique ways to leverage their existing data assets to improve their quality measures.
The payout for the new Home Health Value-Based Purchasing incentives will likely be based on a set of quality performance metrics. These measurements will most likely be derived from OASIS* forms, patient surveys, and claims-based data.
Note: The Outcome and Assessment Information Set (OASIS) is the patient-specific, standardized assessment used in Medicare home health care to plan care, determine reimbursement, and measure quality.
You want your home health agency to be above the median of competing agencies in your state, which gives your agency the most likelihood to collect a program incentive payment. In order to achieve this, you must find new ways to improve quality.
In this article, we will walk you through applying Machine Learning methods, specifically Supervised Learning, to acquire more sophisticated insights into your patient populations. We’ll also cover how to use your data to define and proactively engage populations at risk of not meeting quality standards, an outcome which would hinder you from receiving incentive payments.
What is Supervised Learning?
Supervised Learning is a branch of Machine Learning in which we uncover specific relationship or structures in existing data to effectively produce correct output data. Simply put, we’re looking for trends in our historical datasets. And we can use that information to identify the likelihood of future patient behavior based on those past trends. For our purposes, we will divide Supervised Learning into two contexts: classification or regression**. In this article, we will only focus on classification.
* Classification uses input data relationships and structure in order to predict an output label. And that output label is discrete single or multiple values such that the input values must be classified in to any number of predefined set of planned output labels.
** Regression problem is where we are also using historical data to train our model though the output data produced is a continuous number such as a risk score anywhere from 0 to 1 and is not set as a specific set of predetermined classes as in classification problem.
Supervised Learning, Classification, and OASIS-Based Quality Metrics
In classification, we train our model based on historical data to predict an output label.
So what does that mean in terms of OASIS-based quality metrics?
Let’s look at the quality metric of “improvement in ambulation”, which measures how impaired the level of ambulation is at discharge compared to the start or resumption of care. This measure seeks to quantify how well patients can ambulate, and home health agencies are given credit for compliance on this metric if a patient meets a set criteria.
Home health agencies may have a way to identify patients who have already completed episodes and know what that patient’s pre-episode ambulation level was and if they improved or not. However, you’re already late if you’re waiting until the end of an episode to review ambulation levels in order to perform an intervention. By then, the episode is already at the discharge stage. We can’t go back and update the ambulation level since the moment to give the intervention has passed.
This is where using Supervised Learning comes in. You can predict which patients will be most likely not to meet the criteria for improved ambulation by the end of their episode.
We need to know before an episode ends which patients to focus on. Supervised Learning allows us to stratify all patients based on their likelihood of missing their ambulation goals.
In order to do this, we can use the classification method of Supervised Learning. This would take historical data of patients that shared characteristics with incoming patients (i.e., input data) and determine if the historical data gives us true context into whether a patient will meet the ambulation goal.
The Limitations of Supervised Learning
When the input data used to build your predictive model is the only driver of your output label, you’ll see that there are a few common issues that can arise and must be considered.
First, you simply won’t have enough data to create true learning. Probably one of the most famous quotes defending the power of data is that of Google’s Research Director Peter Norvig, who claimed that “We don’t have better algorithms. We just have more data”. This is the same case with the future of using Machine Learning in Home Health Value-Based Purchasing Quality Measures. The more data you have, the more accurately you can capture relationships among the input data and produce reliable output classifications.
The problem with limited data often leads to a second problem: creating a data model that over-fits the input training data. Meaning, if you have a small number of data points from your training data, your model will strictly learn your training data and won’t be able to learn actual trends and structures in the input data to predict a reliable output.
Where to Go From Here?
“We’re entering a new world in which data may be more important than software.” – Tim O’Reilly, founder, O’Reilly Media.
We’ve entered into a new era where data will be a differentiator most businesses will need to leverage in order to compete. And the home health sector’s ability to compete using its data assets is providing ever more opportunities to excel.
Using Machine Learning and Supervised Learning is just one facet where you can create a competitive advantage for your agency and accelerate your agency’s successful adoption of the value-based purchasing model.