Decision support systems for the treatment of community-acquired pneumonia.

Abstract

Delay to antibiotic treatment of community-acquired pneumonia (CAP) greater than 4 hours following hospital admission is associated with a 15% increase in mortality. Paper-based guidelines have been widely introduced to improve CAP care, but these interventions have under-performed due to poor compliance in complex clinical workflows. Unlike passive paper-based guidelines, alerting systems based on computer-based decision support systems (CDSS) have the capacity to actively draw attention to delayed clinical processes. Formal consideration of local workflow is key to the design and successful implementation of CDSS. I used workflow analysis techniques to develop an evidence-based alerting system designed to reduce the delay to treatment of CAP in the emergency department (ED) of an Australian tertiary hospital. A sample of 6 CAP patients were observed during October 2001 to derive a structural process flow model, which was refined via stakeholder interview. A deterministic process flow model was then developed using an existing retrospectively compiled CAP database, consisting of 246 patients admitted June-December 1998 and 146 patients admitted May-December 2000. A stratified control sample presenting with respiratory symptoms (n=74, January-December 2003) was collected for the assessment of diagnosis and chest x-ray (CXR) accuracy. Treatment delay greater than 4 hours was associated with failure to diagnose CAP in the ED, the absence of CXR evidence, low triage score, delayed CXR, and failure to treat in the ED. ED physicians only identified 54-57% of those discharged with CAP. Radiologists only reported CAP features in 47% - 67% of initial CXRs for these patients. I hypothesised that a CDSS-based alerting system, composed of a CAP early diagnosis model (EDM) and a simple risk model (CRB-65), would identify enough CAP patients to reduce the percentage treated after 4 hours. I constructed an evidence-based naïve Bayesian EDM (sensitivity = 36%, specificity = 93%). It was able to identify 24% of CAP patients that died in hospital, 38% of those with antibiotics delayed greater than 4 hours, and 26% of those with CXR delayed greater than 4 hours. CAP-specific risk models were equivalent to the Australasian Triage Score (ATS) in predicting mortality. I simulated alerting policy by combining the CDSS with the deterministic process flow model. Alerting for treatment at triage or initial physician assessment, when the EDM was positive, approximately halved the median treatment time of 5.53 hours, and decreased the number treated after 4 hours (62%) by 1/3. Treating EDM-positive patients as ATS category 2 produced a similar effect. Current triage practices, embodied mainly by the disease-independent, sign and symptom based ATS are too coarse to deal with conditions such as CAP, where there is high diagnostic uncertainty and delays in diagnosis and treatment are critical determinants of outcomes. Better outcomes may be achieved with quicker diagnostic and treatment workflows via: analysis of current diagnosis and treatment workflows, analysis and correlation of a comprehensive set of patient symptoms, signs and risk factors for the specific disease, and improving triaging and subsequent workflow through a disease-specific CDSS based on early diagnostic models derived from the previous analyses.