From signal detection to verification Marie Lindquist Safety assessment steps Knowledge Benefit > harm Signal verification Signal quantification Signal strengthening Signal detection Effectiveness - risk assessment Evidence of causality and probability Incidence assessment Characterisation of problem Hypothesis generation Time Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - advantages • A unique source of safety signals • Gives information about effects in ’real world’ use – as opposed to the often idealised conditions in e.g. clinical trials • Has a potential to identify rare reactions • Gives indications about which problems are considered most troublesome in practice Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - advantages • Continuous data collection • Low cost • Covers whole populations (at least in theory) • Exists in 90+ countries • All countries have access to each other’s data through the WHO Programme network Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - possibilities • Adverse reaction reports can tell us a lot about the concerns of – treating and prescribing doctors/dentists/midwifes – other health professionals involved in patients’ medicines treatment – patients themselves! Marie Lindquist, Uppsala Monitoring Centre Spontaneous reportingpossibilities • New players – can give more/different information and better data quality • More resources for data analysis – instead of spending all money on perfect reporting systems and bureaucratic processes • Widened horizons – focus on patient safety instead of drug problems Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - problems • Low levels of participation • Bureaucratic systems • Lack of data quality • Causality assessment difficult • Generalisability? • Lack of dialogue and collaboration between stakeholders Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - problems • Not suited for detection of – Delayed reactions – ADRs with a high disease background incidence • Confounding – where disease under treatment could be linked to the outcome • Diagnostic bias • Channeling – Susceptible patients switched to a new drug thought to be better • Underreporting and missing data Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting - problems We have a numerator Frequency of reporting Frequency of exposure But not a denominator! Marie Lindquist, Uppsala Monitoring Centre Epidemiology basics Cases Controls Exposed population A B Non-exposed population C D Frequency Risk Incidence = # cases developing over a defined time period Prevalence = # cases at a given point in time Absolute risk (exp) A/(A+B) Reference risk (non-exp) C/(C+D) Marie Lindquist, Uppsala Monitoring Centre Relative risk Risk (exp)/Risk (non-exp) With case reports alone, we cannot • Calculate incidence/prevalence – Do not know the frequency of the ADR*) – Do not know the size and characteristics of the exposed population • Calculate relative risk – Do not know the frequency of the reaction among the non-exposed – Do not know the size and characteristics of the nonexposed population *) due to (variable) under-reporting Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting or epidemiological studies? • Case report • ? • Case series • ? • Signal => Hypothesis • Hypothesis • Analysis • Study design – Qualitative • Study execution – Reporting frequency • Analysis – Causality – Incidence – Probability Marie Lindquist, Uppsala Monitoring Centre – Frequency – Incidence – Probability – Causality Spontaneous reporting or epidemiological studies? Signal detection Signal Signal strengthening quantification Case reports Signal verification Epidemiology Statistical evidence Clinical & pharmacological information Marie Lindquist, Uppsala Monitoring Centre We need both Spontaneous reporting • Good information on individual harm Epidemiology • Good information on studied population but • Generalisation to whole populations? Marie Lindquist, Uppsala Monitoring Centre • Incidence not translatable to individual risk Signal strenghtening Using spontaneous reports Qualitative signal assessments • Documentation level – Are all relevant facts present? • Credibility of cases – Result and accuracy of individual causality assessment • Credibility of signal – Enough cases? – Pointing in the same direction? – Absence of reverse findings? Marie Lindquist, Uppsala Monitoring Centre Building a signal 4 feasible cases 3 index cases 2 substantial cases 1 index case 7 3 case reports case reports Marie Lindquist, Uppsala Monitoring Centre Qualitative signal assessments • Confounding – Other risk factors than drug present? • Bias – Selection, publication, measurement bias present? • Individual and public health impact of the signal – Degree and duration of harm caused Marie Lindquist, Uppsala Monitoring Centre Quantitative signal detection using surrogate denominators • Use other reports in database – as surrogates for exposure and controls – to calculate statistical disproportionality Marie Lindquist, Uppsala Monitoring Centre Quantitative signal detection methods Cases Exposed population Non-exposed population Controls Drug X+ ADR Y Drug X+ Other ADRs Other drugs + ADR Y Other drugs+ Other ADRs Marie Lindquist, Uppsala Monitoring Centre Signal strengthening Using drug usage denominator data A realistic first approach Make an estimate of the incidence in the exposed population • Use existing ADR data (numerator) AND • Drug use data (denominator) Marie Lindquist, Uppsala Monitoring Centre Possible measures of drug use • How much was actually taken by patients? • How much was dispensed? • How much was prescribed? • How much was sold? Level of certainty of measure Marie Lindquist, Uppsala Monitoring Centre What measures are available? • Actual consumption data • Dispensing data • Prescription data • Sales data Availability of data Marie Lindquist, Uppsala Monitoring Centre Sources of drug usage data • Sales data – IMS (worldwide) – National wholesales information – Hospital/pharmacy information • Prescription data – IMS medical indices (worldwide) – National prescription databases – Healthcare databases • GPRD, Thin/EPIC, Saskatchewan, Medicaid etc. Marie Lindquist, Uppsala Monitoring Centre Sales data as denominator Marie Lindquist, Uppsala Monitoring Centre Combining ADR and sales data Reporting rates are calculated as Number of ADR reports Amount drug sold (exposure surrogate) Marie Lindquist, Uppsala Monitoring Centre Prescription data as denominator Marie Lindquist, Uppsala Monitoring Centre Combining ADR and prescription data Comparable reporting rates are calculated as Number of ADR reports # prescriptions (exposure surrogate) Marie Lindquist, Uppsala Monitoring Centre Reporting rates using sales or prescription data Can give an indication of – The size of the problem (incidence estimate) – Differences between the target drug and its comparator drugs – Differences between countries – Differences over time (if longitudinal data is used) Marie Lindquist, Uppsala Monitoring Centre Reporting rates using prescription data In addition to when sales data is used, analyses can be made involving – Patient age – Gender – Indication – Dosage Marie Lindquist, Uppsala Monitoring Centre Using this method were able to • Identify new problems on old drugs • Elucidate reasons for country differences, e.g. – Differences in drug usage – Publication bias influences reporting • Identify differences in reporting patterns between drugs • Evaluate if an identified signal for one drug may be a group/class effect • Show that the public health impact of a signal could have been recognised earlier Marie Lindquist, Uppsala Monitoring Centre Reports of carbamazepine and Stevens Johnson syndrome in WHO database Country # reports Thailand 318 United States 155 Germany 144 Malaysia 129 United Kingdom 104 Spain 37 Sweden 31 Canada 28 Switzerland 16 Italy 10 Netherlands 8 Denmark 2 Marie Lindquist, Uppsala Monitoring Centre Adding sales data Country Malaysia Thailand Sweden Switzerland United Kingdom Netherlands Germany Spain Canada United States Denmark Italy Marie Lindquist, Uppsala Monitoring Centre # reports sum KG sales sum 129 4,705.9 318 28,797.7 31 95,036.4 16 53,641.7 104 604,275.6 8 53,917.6 144 990,518.6 37 307,818.3 28 258,180.9 155 2,086,026.5 2 44,577.6 10 448,681.8 # reports/ mill DDD 27.4 11.0 0.3 0.3 0.2 0.1 0.1 0.1 0.1 0.1 0.0 0.0 Lessons learned • Direct comparisons drugs/countries should not be made without investigating possible reasons for differences • The choice of comparator drug/s is important – problems if not used for the same indication, or if the severity of the disease treated is different – analyses should be made at similar times in their marketed lives • Biases in denominator and numerator can be different – E.g. under-reporting in numerator and not in denominator Marie Lindquist, Uppsala Monitoring Centre Lessons learned • Method only useful when there is a relatively large amount of reports – particularly when involving more variables than drug/ country/ year • 'Push button' merging of the different data sets is not possible – more or less extensive manual mapping/data washing needed Marie Lindquist, Uppsala Monitoring Centre Adding drug usage denominators conclusion • Combining drug use data and spontaneous ADR data is helpful when analysing signals – But reporting rate ≠ incidence – Gives a more or less good estimate of incidence – Need data on background rate of disease to compare risk in exposed/non-exposed populations • Be careful when making comparisons between drugs/countries/different time periods – be aware of all factors that can influence denominator and numerator Marie Lindquist, Uppsala Monitoring Centre Spontaneous reporting or epidemiological studies? Signal detection Signal Signal strengthening quantification Case reports Signal verification Epidemiology Statistical evidence Clinical & pharmacological information Marie Lindquist, Uppsala Monitoring Centre Towards signal verification • Cohort event monitoring • Signal testing using longitudinal health care data Cases + controls (exposed) • Case-control studies As above + non-exposed cases and controls • Cohort studies • Other study designs • Randomised clinical trials Marie Lindquist, Uppsala Monitoring Centre As above + study population randomly divided into exposed/nonexposed Type Question answered Weaknesses Case reports (spontaneous reporting) Reporting frequency of No controls, no information a particular outcome on exposure Case reports + usage data Reporting rate (number of cases compared with data on how much a drug was used) Marie Lindquist, Uppsala Monitoring Centre Strengths Cheap, covers whole country populations (in theory). Can detect rare reactions, postmarketing Cases and users (exposure Quick, cheap surrogate) not from the same population. No comparison with non-exposed. Gives only a crude estimate of incidence in the exposed Type Case-control study Question answered Strength of relationship outcome - exposure (how many were exposed and how many were not among the cases; and the same for the controls). Starting with identified cases. Can be done retrospectively Cohort study Weaknesses Matching cases and controls Relatively quick and cheap. Can controls not necessarily study multiple exposures. representative of the whole study Works for rare diseases/ADRs population. Risk of bias and confounding. Less certain causality Incidence - how many will Takes time, and high cost. develop the outcome (cases) and Exposure decided on clinical how many will not in the grounds (by study leader -risk of exposed population and in the bias). Sometimes limited in size non-exposed population. and power (to detect rare outcomes) Starting by allocating exposure. Prospective Randomised controlled As with cohort studies, but trial population randomly divided into exposed/non-exposed Marie Lindquist, Uppsala Monitoring Centre Strengths Takes time, and high cost. Practically limited in size and power. Sometimes ethical problems withholding treatment from controls Can look at multiple outcomes. Less risk of bias and confounding than in casecontrol studies. Gives incidence Risk more likely to be correct since possible selection bias can be ruled out (random allocation of exposure), and confounding can be controlled for