Semi-automated surveillance systems save time compared with traditional manual methods, particularly for non-ventilator hospital-acquired pneumonia (nvHAP), a nosocomial infection which can affect all non-intubated patients. In semi-automated surveillance, a computerized algorithm selects patients with high probability (i.e. "at risk") for subsequent manual confirmation. This study aimed to evaluate the performance of several single indicators and algorithms to preselect patients at risk for nvHAP.
Single nvHAP indicators, identified based on literature, expert opinion and data availability, were combined to simple and complex algorithms. Both single indicators and algorithms were applied on a patient cohort of 157 902 patients, including 947 patients with nvHAP according to our reference standard, i.e. validated semi-automated nvHAP surveillance system plus the manual surveillance of patients with hospital-acquired pneumonia discharge diagnostic codes. Performance characteristics like sensitivity, workload reduction, and number of patients needed to be screened to detect one case of nvHAP were assessed.
Compared with the reference standard, single indicators had a sensitivity ranging from 35.1% (332/947) (oxygen desaturation) to 99.7% (944/947) (radiologic procedure). The workload reduction varied from 57.3% (90 505/157 902) (length of hospital stay >5 days) to 98.4% (155 453/157 902) (ICD-10 discharge diagnostic code). The highest workload reduction was found in complex algorithms, e.g. the combination "radiologic procedure including full text AND temporally related abnormal white blood count or fever AND antimicrobials AND C-reactive protein AND decreased oxygenation AND hospital stay ≥5 days AND no intubation" which reduced the number of patients who have to undergo manual review by 96.2% (151 867/157 902), while maintaining a sensitivity of 92% (871/947). The number needed to screen applying this algorithm was 6.4 patients.
Several single indicators and algorithms showed a high workload reduction and a sensitivity above the defined threshold of 90%. Our results could assist hospitals or stakeholders of surveillance initiatives in developing algorithms customized to their local conditions.