Research Programme 3: Pharmacovigilance
Prof Helen Colhoun, Prof Paul McKeigue, Dr Ewan Pearson, Prof Frank Sullivan
Aim
- To link community prescribing data to EPRs to demonstrate the feasibility of national pharmacovigilance
Background
Existing methods for discovery and ascertainment of adverse drug reactions (ADRs) have serious limitations: some are too rare to show up in clinical trials, the spontaneous reporting rate (e.g. for yellow-card systems) is very low and many adverse reactions are not recognized until a drug has been on the market for a long time. SHIS-R in which we will harness the ability in Scotland to routinely link prescription data with morbidity data for a large number of patients has the potential to overcome these limitations. In principle it is possible to use these data to test specific hypotheses about drug-adverse event associations and also to discover previously unsuspected drug effects on morbidity: adverse and beneficial. However, new data analysis methods and tools are needed to fully exploit these data.
The challenges
Finding drug effects from observational morbidity and prescribing data presents several challenges. For example we will need to characterise data quality and improve the metadata held on the key prescription items held in the relevant databases.
Another challenge is that drug-event associations can be confounded by co-morbidity. For effects with short latency comparisons between exposed and unexposed spells (case-crossover) within individuals can help to overcome this. For effects with longer latency having detailed longitudinal data on co-morbidity, such as is held on our clinical databases, is key to minimising residual confounding following covariate adjustment.
We want to be able to search for clustering of drug exposure with morbidity to discover previously unsuspected adverse drug-event associations. This requires tools that can search over a very large space of possible models, penalizing over-complex models that are less probable. In principle, we can use Bayesian methods to do this.
Research question
Can we create analytical tools and approaches that can be used to test hypotheses about adverse drug reactions, and to discover previously-unsuspected effects of drugs on morbidity in longitudinal datasets that link drug exposure and morbid events in a defined population?
Methods
We will use the extensive diabetes clinical database (SCI-DC) with linkage to other databases to develop analytic tools and approaches that could then be applied to other clinical databases, including potentially wholly primary care databases such as QResearch in practices using EMIS. In addition linkage to the SMR and dispensed prescriptions is increasingly possible across Scotland and has been complete within one Health Board (Tayside) since 1993. Linkage to the hospital laboratory biochemistry database will also be carried out.
First we will use these data to establish the evidence for, incidence rates of, and risk factors for specific suspected ADRs for commonly used diabetes and diabetes related drugs including anti-hypertensives. For these analyses we will use established multivariate methods, including survival analysis and multilevel modelling, augmented with newer data analysis tools that deal with missing data. With over 200,000 type 2 diabetic patients each with several years observation we have power to examine a wide range of hypotheses. When we extend the use of our approach beyond diabetes, we will be able to extend these techniques into long-term complications such as investigation of the potential link between photosensitising drugs and cutaneous squamous carcinoma and the hepatotoxicity af a wide range of agents.
Discovery of previously unsuspected adverse drug-event associations can be viewed as an unsupervised learning problem, based on inferring association and temporal sequence in both case-crossover and between-individual comparisons. The relationships of drug exposure and confounders to outcome can be represented as a directed acyclic graphical model (a special case of the more general family of models known as Bayesian networks. The problem is to learn which models (out of all possible models in the prior hypothesis space) are supported by the data. The methods that we propose to develop can be applied to search for clustering of drug exposure with unsuspected adverse events, or to test specific hypotheses of current interest. We will validate these tools by demonstrating that they can detect known ADRs e.g. myalgic and myopathic outcomes with statin therapy.
Generalisability
Although the development of these methods will be undertaken using the SCI-DC database as a test-bed, the approach will be generally applicable to any electronic medical record system that includes both drug exposure data and morbidity. In the longer term these techniques will be widely applicable to UK electronic health records as use of the NHS number is rolled out.