JMIR Publications

ABSTRACT

Background:

Patients develop pressure injuries in the hospital due to long exposure to localized pressure and low mobility, among other predisposing factors such as circulation conditions. More than 2.5 million Americans develop pressure injuries annually. The Center for Medicare and Medicaid considers hospital-acquired pressure injuries as the most frequent preventable event, and they are the 2nd most common claim in lawsuits. With the growing utilization of electronic health records in hospitals, an opportunity exists to explore these records using artificial intelligence to more closely capture hospital-acquired pressure injury conditions rather than relying on occasional manual assessments by human experts. Electronic health records, both structured and unstructured, contain data on hospital-acquired pressure injury (HAPI). Different data sources within a database may provide conflicting information on HAPI. Clinically justified HAPI rate is important for administrative decisions and accurate HAPI labelling is essential for computational models.

Objective:

We identified discrepancies in HAPI sources within electronic records, explored the conflicts and overlaps in information and provided a unified cohort definition for establishing HAPI classification labels. This definition conforms better to clinical guidelines for the HAPI nursing quality metric than existing definitions.

Methods:

Existing data-driven based definitions of hospital-acquired pressure injuries do not align with the clinical guidelines which reduces their adoption and are conflicting even for the same patient population. Moreover, these inconsistent criteria make it impossible to compare the success of computational methods defined through them to identify hospital-acquired pressure injury. In this retrospective cohort study, we analyzed the congruence among HAPI data in clinical notes, diagnosis codes, and medical events from the Medical Information Mart for Intensive Care III (MIMIC-III) database. Based on conflicts among data sources, we analyzed four HAPI cohorts with different criteria. We assessed the cohorts’ performance for HAPI classification using tree-based and sequential neural network classifiers.

Results:

We analyze the complexity of defining hospital-acquired pressure injuries using diverse but inconsistent data sources, provide a definition that more closely resembles nursing guidelines, and showcase the higher accuracy of a hospital-acquired pressure injury prediction model based on our definition on a large dataset. This analysis highlights the importance of standardized HAPI criteria and shows computational advantage of more refined cohort definitions for HAPI classification from patient notes. Our proposed cohort Emory HAPI (EHAPI) performed better than three existing cohorts with a higher average performance on the test set across ten runs for both the tree-based and neural network HAPI classifiers using MIMIC-III data.

Conclusions:

Standardized HAPI definitions are important for accurate HAPI nursing quality metric automation and determining HAPI incidence for preventive measures. EHAPI definition has favorable properties making it a suitable candidate for HAPI classification tasks.