Starting this year, as part of Value-Based Purchasing expansion, the Department of Health and Human Services began rating home health agency performance on several quality metrics (some based on OASIS data) to determine payouts.
In our last article, we discussed Supervised Learning and how it uses ground rules, such as labeled historical data, to predict future values of a quality metric. In our example, the labeled data was historical data about patients who improved their ambulation (performance on a quality metric), as reported on the OASIS assessment form done at patient intake. We used reported quality results from previous patient episodes to turn our data into a tool to predict the likelihood for new episodes to meet the quality criteria at discharge. In our example, Supervised Learning allowed us to use our data to predict which patients may fail to meet ambulation metrics so that we could start interventions earlier.
Unsupervised Learning Methodology
Unsupervised Learning is slightly different in that we do not provide explicit labels. In this instance, we do not really pose a problem based on us trying to predict a classification. We are not trying to capture the likelihood of a new episode (i.e., new input data) meeting a quality measure (i.e., output classification) from historical data.
One way Unsupervised Learning can be used is to perform exploratory analysis of data and find underlying patterns.
Say we want to find trends about a common characteristic of certain patients. We could use clustering of our data to find similar attributes. Then we could possibly influence similar patients who don’t meet specific quality criteria.
We could perform this exploratory analysis on patients who perhaps share, for example, similar socio-economic characteristics. If the home health agency finds that a patient in this group is not following a prescribed care plan, the agency could use Unsupervised Learning to identify characteristics of patients that should be targeted and develop specialized approaches for those patients.
Social or economic determinants of health are a quite common underlying factor. Home health professional try to address these issues for quality improvement. But a human could have difficult gaining the insight from massive amounts of data needed to propose trends.
This is where Unsupervised Learning comes in. Using clustering can help agencies find patterns and trends among patients who are not meeting quality metrics and give management insight to develop new insights into patient behavior.
This is only one of surely many trends in your data that a human may miss. Initial insights simply might not be so straightforward. With Unsupervised Learning, one can take those initial insights and test an individual hypothesis based on the findings.
Limitations of Unsupervised Learning
One limitation of Unsupervised Learning is the fact that, since we don’t have any labels telling us a ground truth, it is hard to determine the model performance of unsupervised approaches.
As you might recall in the Supervised Learning example, we had an initial set of patients and at the end of the episode, we could tell if those patients showed improvement in ambulation, the quality metric we were trying to classify based on historical data. At the end of the cycle, we knew if our model met reality or fell short.
In Unsupervised Learning, the goal and objective is purely to find structure in the underlying data, which means results can be more subject to interpretation.