Improving Patient Outcomes by Associating Comorbidity Relationships

With the onset of Covid-19, the word “comorbidity” has been frequently mentioned in articles, social media, and news programs. If you have been tuning in over the past two years, you know that some comorbidities dramatically increase a person’s risk of severe side effects from Covid. The health implications of comorbidities, however, reach far beyond their Covid implications. Comorbidities are a leading indicator of a person’s general health. Acquiring multiple comorbidities can have devastating impacts on a person’s health and cost of care. Understanding the risk that a patient has of acquiring multiple comorbidities allows health care providers to create preventative care plans for their patients to improve their health outcomes. 

So, what are comorbidities?

Comorbidities are chronic medical conditions known to be associated with or lead to other chronic medical conditions. For example, patients with Type II Diabetes tend to also have hypertension, which can also contribute to additional health complications. In addition, patients with comorbidities tend to be costlier to care for and have a higher risk of developing additional underlying health problems. Through data mining and machine learning techniques, Zencos uses analytics to identify patients at risk of acquiring new comorbidities, thus allowing health care providers to create preventative care strategies for their at-risk patients.

Machine learning and health care have a well-established relationship. In the big wide world of data science, there are a plethora of fancy-pants algorithms available for analyzing health care data – each more titillating than the next. With names like Support Vector Machine, eXtreme Gradient Boosting (XGboost), and Long Short-Term Memory Recurrent Neural Network, it’s impossible not to get excited by the possibilities. One could spend weeks getting lost in the proverbial rabbit hole of empirical studies showing how data mining and machine learning can be used to improve patient outcomes. To understand the relationships that exist between comorbidities, there is one relatively simple algorithm that works especially well – association analysis.

Association Analytics in the Real World

Association analysis (specifically the Apriori algorithm) first found its footing in the retail world. Association analysis was specifically designed to find what might otherwise be hidden relationships between two or more products. The idea is to take transactional data across many customers and identify what items are frequently purchased together. Today, however, association analysis has moved far beyond traditional customer transaction datasets. As a result, you can find this “market basket” type of analysis being used in a variety of industries, with examples listed below.

–          Entertainment: Association analysis is part of the process that services like Spotify and Netflix use to recommend your next great entertainment experience

–          Accademia: Researchers often use association analysis for text classification to identify overall messages or sentiment

–          Social Media: Social media platforms use association analysis to try to figure out who your next best friend is going to be

Health care is especially poised to take advantage of the data mining opportunities afforded through association analysis. Whether identifying patterns in oral cancer detection, predicting heart disease from medical records, or – for the purposes of this blog – detecting comorbidity relationships from patient medical records, association analysis is a go-to data mining solution driving improved patient outcomes.  Analysts can use association analysis to leverage years of patient medical claims to identify patterns that would otherwise be undetectable.

Rules to Live By

Unlike predictive models where you return an actual prediction (either a probability of an event or a predicted value of a target variable), association analysis identifies rules. The rule has two sides: the left side (the antecedent) and the right side (the consequent). Association rules are typically written in {A,B} = > {C} format. Or for my English speakers out there, if A and B, then C.

Once the analysis has been done, and the rules are derived, association analysis provides three metrics to evaluate the quality of a given rule:

1. Support: This is a measure of how frequently the items, or in our case, comorbidities in a rule appear within the population being measured.
2. Confidence: Confidence is a conditional probability for the rule being presented. This gives the frequency in which the comorbidity on the right side is observed, given that the comorbidity listed on the left side of the rule is observed.
3.  Lift: Lift tells us how much more likely a patient has the comorbidity on the right side of the rule when the comorbidities on the left side of the rule are present than a person is to have the comorbidity on the right side of the rule at random.

In practical terms, support, confidence, and lift work together to determine just how impactful a rule may be. Let’s say we ran our analysis and identify the following rule:

{Hypertension Complicated (I.11)} => {Renal Failure (N18.1-N18.3)}   

The rule has a support of 0.08, confidence of 0.4, and a lift of 19.5. The interpretation of this rule would be as follows:

• The combination of Complicated Hypertension and Renal Failure show up in 8 percent of our patient population
• Of the members that have Complicated Hypertension, 40 percent also have Renal Failure
• And finally, someone with Complicated Hypertension is 19.5 times more likely to have Renal Failure than someone is to have Renal Failure at random.

Association analysis rules provide clear and actionable insights for health care providers. Once a quality rule is identified, the first step is to screen all of the health care provider’s patients who have the antecedent of the given rule but do not have the consequent. For example, after identifying that a member with Complicated Hypertension is 19 times more likely to have Renal Failure, a provider should screen a patient with Complicated Hypertension for Renal Failure to make sure they are not living with an unidentified condition. If the patient does not have the consequent (in this example, Renal Failure), the patient’s primary care provider should create preventative care strategies to keep the patient from acquiring the additional comorbidity in the future.

Association Analysis for Good

Zencos frequently applies association analysis to medical records to identify patterns, or “rules”, in our customers’ data. Our customers are then able to use the identified patterns to improve health outcomes of their patient population. Though identifying these underlying relationships in comorbidities is just one example of the application of association analysis in health care, it can immediately impact both the cost and quality of care for patients.