Using Visual Analytic Methods to Identify Patient Groups

• Download Summary

What was the project about?

Patients with multiple health problems, such as diabetes and heart disease, may benefit from personalized treatment approaches. For example, patients who have both breathing problems and diabetes may need different medicines than patients who have diabetes alone. Researchers can use statistical methods to group patients with specific health problems and figure out how well treatments work for those patients. But current methods don’t always find all the health problems that can affect how treatments work.

In this project, the research team developed a new method to group patients with common health problems. The method, called a visual analytic method, used a computer program and patient data to draw pictures or maps of the patient groups. The method helped the research team figure out the chance of returning to the hospital for patients in each group.

What did the research team do?

The research team used Medicare claims data for patients with COPD, a lung problem that makes it hard to breathe. The team looked at data for 29,016 patients who returned to the hospital within 30 days of discharge. Then they looked at data for 29,016 patients who didn’t return to the hospital within 90 days of discharge. The team matched patients who did and didn’t return to the hospital based on their age, gender, race, and income.

The research team then used the new method to identify and group patients who returned to the hospital, based on their health problems. The method created a picture showing how the groups were similar and different. Then the team used statistical methods to figure out the chance of returning to the hospital for patients in each group. The method also showed a patient’s likelihood of being in a specific group.

Doctors helped the research team develop the methods and review the results.

What were the results?

The method displayed a picture of four groups of patients with COPD based on their most common health problems:

• Patients with high blood pressure
• Patients with diabetes with complications, kidney failure, and heart failure
• Patients with mental health problems and social concerns
• Patients with organ damage and digestive conditions

The method also identified the chance of returning to the hospital for each group. For example, the patients with diabetes group had an 18 percent chance of returning to the hospital.

What were the limits of the project?

The new method took almost a week for a computer to run. The research team used claims data. Results may have differed if the team used other types of data or selected patients in different ways.

Future studies could test these methods using other types of data such as health records.

How can people use the results?

Researchers could use these methods to identify groups of patients with certain health conditions, and figure out how treatments work for patients in those groups.

Background

Patient comorbidities affect how patients experience illness and respond to treatments. Distinguishing heterogeneous subgroups of patients with specific comorbidities may help explain disease progression and treatment response for patients in these subgroups. Current methods for identifying heterogeneity group together patients with common characteristics; however, these methods may not capture all common comorbidities or reveal relationships among different patient subgroups. New methods that identify clinically important relationships within and across subgroups may improve understanding of heterogeneity and help clinicians and patients make personalized treatment decisions.

Objective

To develop a visual analytic method for identifying subgroups of patients with COPD readmitted to the hospital, based on their comorbidities

Study Design

Methods

Using Medicare claims, researchers identified patients with COPD with a hospital readmission within 30 days after discharge. They matched 29,016 patients based on age, gender, race, and income to a control group of patients with COPD who were not readmitted within 90 days. Researchers randomly divided the matched data into two sets of 14,508 case-control pairs: a training data set and a replication data set.

Researchers trained a visual analytics model with the training data set using a technique called bipartite network analysis. They applied computer algorithms to create a picture of associations between patients and their comorbidities. This picture displayed a network of circles, or nodes, representing patients and their comorbidities, and lines connecting the nodes, which showed relationships between them. Nodes with strong similarities formed visual clusters of subgroups. Based on the visual output, researchers identified subgroup characteristics. Researchers then used the replication data set to test the visual subgroup clusters for significance and replicability. They also calculated each subgroup’s readmission risk.

Using both data sets, researchers used classification modeling to determine a patient’s subgroup membership and confirm results from visual analytic modeling. Multinomial logistic regression predicted the probability of a patient belonging to each subgroup.

Researchers worked with clinicians to develop the analytic approach and interpret results.

Results

Visual analytics modeling produced four statistically and clinically significant subgroups of readmitted patients with COPD with their most common comorbidities (p<0.05) and readmission risk:

• Subgroup 1 had low disease burden, uncomplicated hypertension, and 12.7% risk
• Subgroup 2 had diabetes with complications, renal failure, heart failure, and 17.8% risk
• Subgroup 3 had psychosocial comorbidities and 15.9% risk
• Subgroup 4 had end organ damage including gastrointestinal disorders and 19.6% risk

Classification modeling had a high level of accuracy (99.1% to 100%) and correctly predicted subgroup membership for 99.9% of patients.

Limitations

Visual analytics modeling has high computational burden. When visuals have dense connections, the method may not distinguish relationships between subgroups. Researchers selected cases and controls in different ways, which may have introduced selection bias and affected results.

Conclusions and Relevance

Visual analytic modeling identified clinically meaningful subgroups of patients with COPD and their readmission risk.

Researchers could use these methods to better understand disease progression and treatment outcomes within patient subgroups.

Future Research Needs

Researchers could test the approach using other variables and other types of data such as health records.