Methods: Using the University of Virginia (UVA) electronic health record (EHR) data, we identified all neutropenic patients ≥18 years who were admitted to inpatient wards over a five-year period. Neutropenia was defined as an absolute neutrophil count <500 cells/mm3. Adverse outcomes were defined as 1) development of fever 2) ICU transfer or 3) mortality during hospital stay. We then used Bayesian model averaging to build predictive models for fever and mortality utilizing readily-available clinical laboratory and vital sign data. Finally, we built a combined adversity model (fever and/or ICU transfer and/or mortality) in the same cohort. All models were valideated using a 60:40 testing/training method.
Results: We identified 1056 patients with neutropenia admitted to UVA from 2010 to 2015. Fever occurred in 427 (40.4%) patients and 95 (8.9%) patients died. The final neutropenic fever model had a c-statistic of 0.75 and included the following variables: temperature, pulse rate, white blood cell count, and platelets. The model identified patients a median of 9.6 hours before onset of fever. The final neutropenic mortality model had a c-statistic of 0.79 and included the following variables: sodium, creatinine, potassium, ALT, platelets, albumin, and magnesium. The combined adversity model had a c-statistic of 0.69 and included the following variables: pulse, temperature, respiration, platelets, white blood cells, and albumin. The model identified patients a median of 9.3 hours before onset of fever.
Conclusion: We created predictive models for fever, mortality, and adverse outcomes in neutropenic patients. These models could easily be implemented into the EHR as an EWS to facilitate early intervention and improved outcomes in the neutropenic patient population.
J. Ainsworth, None
P. Rayapati, None
L. Barnes, None
C. Moore, None