Can data from an electronic medical record identify which patients with pneumonia have Pneumocystis carinii infection?

Journal Name: 
Int J Med Inf
BACKGROUND: Pneumocystis carinii is the leading opportunistic pulmonary infection in HIV-infected patients. Invasive diagnostic procedures might be avoided if available electronic data can accurately identify patients with Pneumocystis pneumonia (PCP). METHODS: We extracted data from electronic hospital records, emergency department records, and a pathology database for 299 HIV-infected patients with pneumonia who underwent bronchoscopy. We identified independent indicators of confirmed PCP using logistic regression analysis on a random half of the patients and validated the predictive power of the resulting model on the other half. RESULTS: Bronchoscopy confirmed pneumocystis carinii in 111 patients (37%). Five of the seven significant independent predictors of PCP came from patients' electronic medical records: infiltrate on chest radiograph, male gender, lower red cell distribution width, lower serum creatinine, and a prior positive HIV test. The other two (duration of illness and presence of dyspnea) came from the emergency department record. A simple index found 43% of patients at low risk (18% with pneumocystis), 37% at moderate risk (36% with pneumocystis), and 20% at high risk (74% with pneumocystis). CONCLUSIONS: Data from electronic medical records can help quantify the risk of PCP among HIV-infected patients. However, the model failed to identify 18% of patients with PCP in the low risk group, and empiric therapy would erroneously treat 26% of patients classified as high risk. Bronchoscopy is needed to accurately diagnose PCP among HIV-infected patients with pneumonia. However, if bronchoscopy is not available, the model can help with initial decisions about antibiotic therapy.
analysis, Bronchoscopy, decision, diagnostic, Dyspnea, electronic, Emergencies, hospital, Hospital Records, Male, medical, Medical Records, Methods, pathology, patient, Patients, Record, Records, regression, Regression Analysis, ResNet, Risk, therapy