ACCESSPOINT • VOLUME 5 • ISSUE 10 PAGE 35

The author

David Ansell, mb, CHb, mRCS, PHD is Associate Director, RWE Solutions, IMS HealthDansell@uk.imshealth.com

Prospective identification of drug safety signals from primary care EMR

At a time of growing demand for more accurate and timely drug safety evidence, a landmark study supports the value ofelectronic medical records (EmR) for detecting new adversereactions. It also shows that statistical associations in EmR mustbe treated with as much caution as those from individual casereports − and be subjected to clinical and epidemiologicalreview. A deep understanding of the methodologies, datacollection and clinical practice involved is implicit.

PAGE 36 IMS HEALTH REAL-WORLD EVIDENCE SOLUTIONS

INSIGHTS HEOR, PHARMACOEPIDEMIOLOGY & DRUG SAFETY

Insights from The Health Improvement Network (THIN) databaseIncreasingly stringent regulatory requirementsfor pharmaceutical risk management andsafety surveillance have accelerated researchto improve the detection of new adverse drugreactions (ADRs) under conditions of normalproduct use. For many years, the process of identifying potential signalsand the existence of previously unknown risk has reliedmainly on individual case safety reports (ICSRs). Morerecently the use of longitudinal health data (LHD) has beenexplored, both to complement ICSR information andovercome some inherent limitations. Most studies lookingto apply LHD have investigated its ability to distinguishestablished ADRs from unrelated adverse events; few haveattempted to examine a role for this data in detectingemerging safety signals.

leveraging EmR in pharmacovigilanceMarking an important milestone in efforts to apply EMR inday-to-day pharmacovigilance, a new study has sought toevaluate a process for assessing temporally associated drugsand medical events (adverse events) in this data.1

Specifically, the researchers aimed to determine (1) to whatextent exploratory analysis of EMR would identifyimportant potential safety signals and (2) what proportionof false alarms could be expected if the temporalassociations were taken at face value rather than subjectedto epidemiological review.

Utilizing the Uppsala Monitoring Centre’s vigiTrace™framework for health data exploration, the studycomprised integrating the vigiTrace™ software frameworkwith the primary care EMR and performing a structuredassessment of more than 500 pairs of drugs and medicalevents in THIN (The Health Improvement Network) – anelectronic medical records resource from primary care inthe UK. THIN includes more than 12 million patients, withover 3.8 million being currently active patients. The EMRare collected from general practices and are representativeof the entire UK population in terms of age, gender,medical conditions and death rates. The data extract for thecurrent evaluation was from January 2011 and covered 7.7million patients.

A key element of vigiTrace™ is a graphical display(chronograph) which summarizes and visualizes temporalassociations between two events. In this case, thechronograph focused on the cohort of patients with newprescriptions (Rx) of the drug in question and exploredvariation over time in the recording of a medical eventrelative to those new Rx. Further, it contrasted the observednumber of patients with a record of the particular event toan expected value in each time period, based on an externalcontrol group. VigiTrace™ also provided analytics tosupport the structured assessment with a calibrated self-controlled cohort analysis.

Evaluation process

Over 40 drugs were randomly selected from THIN on thebasis of specified inclusion/exclusion criteria, key amongstwhich was the presence of more than 5,000 new Rx. Medicalevents (up to 20 per drug) chosen at random from thoseidentified as temporally associated with a new Rx of thedrug in question, were assessed for relevance prior toundergoing in-depth analysis.

The in-depth assessment was based on a structuredquestionnaire and included a review of the UK Summary ofProduct Characteristics (SPC) document as well asadditional exploration of data in THIN. Among factorsaddressed as part of this appraisal were: the nature of thetemporal pattern; demographics of the cohort; use ofconcomitant medicines; previous signs and symptoms; andpotential confounding by underlying disease.

Results

From the more than 500 relevant drug-event combinationsthat were identified, 25% were categorized as known ADRs,based on the SPC review (eg, sleep disturbance for a drugwith insomnia listed, glaucoma for a drug with acuteglaucoma listed).

Close to 100 of the remaining combinations were classifiedas meriting full clinical review, beyond the restricted scopeof the study assessment. Examples include multiple organfailure with a selective serotonin reuptake inhibitor (SSRI);skin sensation disturbance (eg, paresthesia, numbness,tingling) with a long-acting beta-2 agonist; and anophthalmic condition with a diuretic. The strength ofevidence for these combinations varied: most of themmerely lacked alternative explanations to suggestivetemporal patterns, whereas a few also had support inexperimental evidence or regulatory information fromother countries than the UK.

In contrast, the majority (approaching 300) of thehighlighted drug-event pairs were deemed unlikely toreflect direct causal relations and hence dismissed fromfurther review. The most common reasons for this wereconfounding by the underlying disease or earlier signs andsymptoms indicating that the onset of the medical eventpreceded the start of drug treatment. Examples includeendometriosis with a drug for the relief of IBS where theprior diagnosis of IBS (based on abdominal pain) was latershown to be endometriosis, and eustacian tube dysfunctionwith antibiotic drops utilized for treating an ear infection.

Examples of the chronograph outputs are shown in Figure 1 opposite.

Prospective identification of drug safety signals from primary care EmR

ACCESSPOINT • VOLUME 5 • ISSUE 10 PAGE 37

That said, the fact that 76% of the drug-event pairs weredismissed from further evaluation following initial review,indicates that signal detection using LHD should form partof a wider, comprehensive process of detailed clinical andepidemiological review. This is an important area for furtherresearch to inform the future role of LHD in signaldetection. It would include examination of individualpatient histories, evaluation of more detailed information inTHIN (eg, temporal patterns for similar drugs and medicalevents) as well as exploration of alternative, complementaryinformation sources such as the scientific literature andcollections of individual case reports. A broader, contextualunderstanding of the methodologies employed, approachesto data collection, and the prevailing medical practice in thesetting being studied would be a key part of this process.

Implications

With this study has come a clear demonstration thatexploratory analysis of EMR is a valid and feasible approachfor detecting important drug safety signals. If a primary careEMR such as THIN is utilized as the source, then signaldetection will be confined to those drugs prescribed withinthe primary care setting. The initial epidemiological reviewrevealed a considerable number of temporally associateddrugs and medical events, ranging from significant, life-threatening conditions to less serious but potentiallyproblematic events for patients. Importantly, some of thesewere conditions that the current pharmacovigilance system,with its reliance on individual ADR case reporting, may bechallenged to capture. Some of the identified events havenot been previously linked to these therapeutic agents andhave highlighted the requirement for further investigation.

1 Cederholm S, Hill G, Asiimwe A, Bate A, Bhayat F, Persson Brobert G, Bergvall T, Ansell D, Star K, Norén GN. Structured assessment forprospective identification of safety signals in electronic medical records: Evaluation in the Health Improvement Network. Drug Saf, 2015; 38:87-100

Figure 1: Sample outputs from the vigiTrace™ chronograph

The study referenced in this article was performed in collaboration with scientists from Eli Lilly, Pfizer, Takeda, Bayer and Cegedim UK,within the public-private partnership PROTECT, which is funded through the European Innovative Medicines Initiative.

509Temporallyassociated drug-event

pairs

382New

127 Already known 291 Dismissed

91 Merit further evaluation

25%

75%

76%

24%

Source: Cederholm S, Asiimwe A, bate A, bhayat F, brobert G, Hill G, Star K, Norén GN. Structured assessment for prospective identification of potentialsafety signals in electronic health records (Poster). 30th International Conference on Pharmacoepidemiology and Therapeutic Risk management (ICPE) 24-27 October, 2014 Taipei, Taiwan

Cannot sleep – insomnia is temporally associated with newprescriptions of reboxetine and was classified as alreadyknown since insomnia is listed as a very common adversereaction to reboxetine in the UK SPC.

Endometriosis is temporally associated with newprescriptions of hyoscine but was dismissed from furtherreview on account of suspected protopathic bias. Hyoscineis given to treat abdominal cramps, which are a commonsymptom of endometriosis.

Epiphora is temporally associated with new prescriptions ofamiloride and was classified as meriting further review onaccount of the suggestive temporal pattern.