Opportunities to reduce delay to antibiotic in community acquired pneumonia: early diagnosis modelling and simulation

Abstract

To assess the accuracy and simulate the impact of an evidence-based early diagnostic support for community acquired pneumonia (CAP), we developed an early diagnostic model (EDM). This model used findings from history and examination and was based on the odds ratio form of Bayes’ rule. Likelihood ratios (LRs) for CAP predictor variables were derived from literature review. The model was validated in a dataset of 146 CAP positive patients, identified by ICD9 discharge diagnoses, and 74 CAP negative controls, who had a respiratory presenting complaint. All patients were aged over 65 years. The EDM was able to identify 36% of CAP patients with a sensitivity of 93% at a threshold probability of 0.3. The model was poorly calibrated, given that the majority of cases had a CAP probability of less than 0.2. Using simulation we found that up to 38% of delayed antibiotics, 26% of delayed chest xrays and 24% of patients who died could be identified using our CAP EDM. The model was validated by comparison to an optimal model using LRs derived from the local dataset and found to be equivalent (area under the receiver operating curve of 0.8 versus 0.85 respectively). The literature model had superior sensitivity (36% versus 27%). The effect of dependence amongst commonly occurring variables was deemed to be minimal given that joint LRs had similar effects on post-test probability in comparison to combined LRs. These findings confirm that despite poor sensitivity and calibration, computer-based EDMs have the capability to reduce time to antibiotic for a sizable proportion of CAP patients and may have an impact on mortality.