Abstract
Community acquired pneumonia (CAP) is one of the most common infectious conditions managed in primary care. CAP may complicate simple respiratory tract infections (RTIs) but there is limited evidence available to inform general practitioners (GPs)of the characteristics of RTI patients who may be at the risk of pneumonia. The purpose of this thesis was to analyse the clinical profile of adult RTI patients to identify variables associated with development of pneumonia within 30 days through a prediction modelling study.The thesis reports a series of inter-related research studies that analysed electronic health records data from the Clinical Practice Research Datalink (CPRD). The research addressed four inter-related objectives. The wider antimicrobial stewardship context was explored through an analysis of antibiotic prescribing records from English general practices participating CPRD from 2014 to 2017. Antibiotic prescriptions for the main groups of common infections managed in the community including respiratory infections, genito-urinary infections (GUTI), infectious skin conditions, eye infections were evaluated. An annual relative reduction rate (RRR) of 6.9% for total antibiotic prescription was detected during the four-year period in the English primary care. Respiratory conditions remained to be the most frequent indications for antibiotic prescriptions among informatively coded consultations, also showed the greatest reduction in prescription rates.
Next, secular trends in the incidence of clinically-diagnosed CAP, clinicallysuspected CAP, influenza and pleural infections were evaluated using CPRD data from 2002 to 2017. Clinically-diagnosed CAP incidence was found to increase over time with an accelerated trend after 2010. For clinically-suspected CAP, an overall contemporaneous trend with an average increasing rate being 3.8% from 2002 to 2008 whereas a faster decline rate of 4.9% thereafter until 2017. Study results together with previous research findings suggested that antibiotic prescribing practice and clinically coding behaviour partly contributed to the apparent increase in clinically-diagnosed CAP in primary care settings.
A systematic review of current evidence of prognostic factors for CAP was conducted to identify candidate predictors for the prediction modelling study. 33 prognostic factors for CAP were identified which could be categorized into six groups: patients’ demographic characteristics, lifestyle, environmental exposures, health conditions, medication prescriptions, disease prevention interventions, clinical management and clinical investigations.
Based on previous study findings, prediction modelling study was conducted with an inclusive approach for possible candidate predictors generated from CPRD data from 2002 to 2017. Analysis included 108,842 patients who consulted for RTIs of whom 16,289 patients re-consulted with pneumonia within 30 days after the RTI index date. Data were analysed using machine learning algorithms for variable selection. Variable selection employed and compared random forest, simple logistic regression, and penalized regression models (Lasso, Ridge and Elastic net). Prediction models were developed using the classification and regression tree (CART) approach, as well as simple logistic regression. Internal and temporal validation were performed. Older age, comorbidity and initial presentation with lower respiratory tract infections (LRTIs) were identified as the main predictors of pneumonia diagnosis. Among patients presented with LRTIs, patients older than 85 remained at higher risk of pneumonia re-consultation despite antibiotic prescriptions were offered; those age between 76 and 85 with two or more comorbidities risk of pneumonia re-consultation persisted even if antibiotic prescriptions were issued. LRTI patients younger than 65 without asthma drug or immunosuppressants treatments appeared to have higher risk of pneumonia re-consultation if clinical discretion did not lead to antibiotic treatment. However, cautions are needed when interpreting such counter-intuitive findings as allocation to antibiotic treatment as well as other disease management procedures were not randomized and confounding by disease indications. Therefore, disease pattern identified among LRTI patients indicated that more attention should be paid to subgroup of LRTI patients to investigate the underlying reasons of primary onset of clinical conditions.
Machine learning techniques may allow the identification of novel disease pattern comparing to conventional modelling approaches, which could be deployed to generate research hypothesis, individualized research designs for inventory clinical trials or provide insights for health policy development.
| Date of Award | 1 Apr 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Martin Gulliford (Supervisor) & Abdel Douiri (Supervisor) |