Background Hospital-acquired pressure ulcers (HAPUs) are costly to treat and can result in Medicare reimbursement penalties. Statistical models can identify patients at greatest HAPU risk and improve prevention.

Objectives To use electronic health record (EHR) data to predict HAPUs among hospitalized patients.

Methods EHR data were obtained from an academic medical center that included hospitalized patients with at least 1 skin examination between 2011–2014. These data contained encounter-level demographic variables, diagnoses, prescription drugs and provider orders. HAPUs were defined by stages III, IV or unstageable pressure ulcers not present-on-admission as a secondary diagnosis, and excluded diagnosis of paraplegia/quadriplegia. Random forests and k-means clustering were applied to reduce the dimensionality of the large dataset. A 2-level mixed-effects logistic regression of patient-encounters evaluated associations between covariates and HAPU incidence (Equation 1).

Results The approach produced a sample population of 23,054 patients with 1,549 HAPUs. The mixed-effects model predicted HAPUs with exceptional (99%) accuracy for a rare event (table 1). The greatest odds ratio (OR) of HAPU incidence was among patients diagnosed with spinal cord injury (ICD-9 907.2: OR=247.4; P<0.001). Other high ORs included osteomyelitis (ICD-9 730: OR=27.7, P<0.001), bed confinement (ICD-9 V49.84: OR=31.7, P<0.001), and prescribed topical/subcutaneous enzymes (OR=5.7, P<0.001).

Conclusions Early detection of HAPUs is feasible and the results of these statistical predictions can allow providers to better target prevention to specific patients. This model also implicates spinal cord injury as a potential risk-factor for unavoidable HAPUs. Providers may be missing opportunities to co-diagnose spinal cord injury with paraplegia/quadriplegia which could improve hospital performance measures.

Equation 1. Mixed-effects Logistic Regression Model

Level-1: Encounter-level Fixed Effects

Level-2: Patient/Cluster-level Random Effect

Where···

i: Patient   j: Encounter⇓