Statistical Methods of Signal Detection

Statistical Methods of Signal Detection

INTRODUCTION

The term ‘Signal Recognition’ arises from electronic engineering, where with radio or radar waves there is a real signal that exists but it is accompanied by ‘noise’ in the background, and there is a need to detect the signal, distinguishing it from the background. This terminology has been used in other contexts, notably in medical diagnosis where similarities to the prob-lems in electronics can also be seen. The terminology of electronics has been continued with ‘Receiver-Operating-Characteristic’ curves. These illustrate that with a given amount of information there must always be a trade-off between the risk of the two different errors of classification: calling noise a signal (a false positive) and calling a true signal noise (a false negative). The sensitivity of a diagnostic test is high when there is a low false negative rate; the specificity of a diagnostic test is high when there is a low false positive rate.

With adverse drug reactions (ADRs) there are two levels of diagnosis of causality: first, diagnosis at a single case level; secondly, at a public health or epidemiological level. ADR causality in an individual patient is not the subject of this chapter, but statistical approaches may help with single cases. The public health and epidemiological perspective is of greatest importance, and statistical methods can be of some help. The objective is to find those signals that are indicative of causal effects, and to reject those signals of effects that are not caused by a particular drug. Where they are of public health significance they will either affect large numbers of individuals or have extremely serious effects in smaller numbers. In these circumstances, the public health view requires that true reactions caused by a medicine be recognised as early as possible. At the same time, those suspected reactions that are not caused by a medicine should be recognised as such and minimal resource should be spent on investigating them.

Signals of potential harmful effects may arise from literature reports, observational epidemiological studies, randomised trials and spontaneous reports of suspected ADRs. In some countries the emphasis is on suspected reactions but in others the emphasis is on adverse events. This chapter will concentrate on the analysis of large volumes of these spontaneous reports. Their source will usually be health profession-als but may also include patients. The early evidence from spontaneous reports can be regarded as a poten-tial ‘signal’. This has been defined as showing a ‘possible causal relationship between an adverse event and a drug. Unknown previously’ (Wood, Coulsen and Eccles, 1994). The object is to distinguish the real signals from ‘noise’ precisely.

The details of spontaneous reporting will not be covered here. The salient feature is that health profes-sionals, particularly doctors, report suspected ADRs centrally; this can be to a regulatory authority or to a company. These reports are processed and entered on to a database. Whether they are reported as suspected ADRs or as adverse events there will inevitably be some background reports that are not caused by the drug. There will often be a very large number of reports, and an essential task is to prioritise those that should be investigated first. The purpose of collecting these reports is to detect signals. Even in countries where reporting of ADRs is supposed to be compulsory, reporting rates will usually be much less than 100%. A typical figure is said to be 10%, but it depends very much on the seriousness and newness of the ADR. In the case of fibrosing colonopathy caused by high-strength pancreatic enzymes, the rate was shown to be 100%.