Model identifies risk factors for early hospital readmission after PCI

A prediction model developed by researchers in Boston found that anticoagulation treatment, frequency of emergency department visits before the procedure and anxiety were associated with early hospital readmission after patients underwent PCI.

Lead researcher Jason H. Wasfy, MD, of Massachusetts General Hospital in Boston, and colleagues published their results in Circulation: Cardiovascular Quality and Outcomes on Aug. 18.

Each year, readmissions after PCIs cost Medicare $359 million, according to the researchers. Although the federal government is penalizing hospitals for readmissions, the researchers noted that registry-derived risk prediction models are limited by modest discrimination and cannot account for unstructured data.

In this model, they extracted unstructured and structured data from electronic health records by using a homegrown text search program. They then manually reviewed the data.

The researchers examined the following variables to determine which ones may be associated with an increased risk of short-term readmission after PCI: patient history of nonadherence, past need for medical interpreter, number of emergency department visits in the past year, a history of anticoagulation, a history of atrial fibrillation or atrial flutter, a history of homelessness, a history of end-stage liver disease, a history of malignancy or any subjective description of patient anxiety by providers.

They mentioned those variables had not been used in previous PCI readmission risk models, but they had been associated with readmission for other conditions.

Patients in the study had participated in the Partners PCI Readmission Project, a database of patients who were readmitted within 30 days of PCI within the Partners Healthcare system in Massachusetts. The PCIs were performed between 2007 and 2011 at Massachusetts General Hospital or Brigham and Women’s Hospital.

In the model, the researchers matched readmitted to non-readmitted patients in a 1:2 ratio based on their predicted risk of 30-day readmission. They identified 888 readmitted patients and 1,776 non-readmitted patients. The predicted risk of readmission was 16.3 percent for readmitted patients and 16.2 percent for non-readmitted patients.

The groups were significantly different with regards to the proportions of patients requiring a medical interpreter (7.9 percent in the readmitted group vs. 5.3 percent in the non-readmitted group), the number of emergency department visits in the past year (1.12 vs. 0.77), history of homelessness (3.2 percent vs. 1.6 percent), history of anticoagulation (33.9 percent vs. 22.1 percent), history of atrial fibrillation or atrial flutter (32.7 percent vs. 28.9 percent), history of presyncope or syncope (27.8 percent vs. 21.3 percent) and history of clinician describing the patient as anxious (69.4 percent vs. 62.4 percent).

After performing a multivariable logistic regression analysis, patients were significantly more likely to be admitted if they had more emergency department visits, received anticoagulation and had a history of being described as anxious.  

“Although this effort does not allow us to create a fully automated [natural language processing] system that could extract information without manual review, this does allow us in principle to establish that manual record review could be facilitated by automated techniques, increasing the quantity of reviewed records,” the researchers wrote. “This work also shows that some of these variables are associated with PCI readmission over and above a registry-derived risk score. This work helps establish the feasibility of using unstructured data in prediction and ultimately is the first step to establishing a fully automated [natural language processing] system to predict readmission risk.”

The study had a few limitations, according to the researchers. They only considered readmissions to the hospital at which the PCI was performed and excluded readmissions to other hospital. They also noted that there are a low number of uninsured patients in eastern Massachusetts, so the study may not be generalizable to other regions of the U.S. In addition, they limited their analysis to prespecified variables.

“We think associations between patient characteristics should be regarded as hypothesis-generating, and we have not been able to show that interventions to reduce these risk markers for readmission in fact lead to lower readmission rates,” they wrote. “Such hypotheses should be tested in prospective trials.”