METHODS

We used data collected by self completion questionnaire from four populations, in three countries (United Kingdom, Australia, and United States) as part of seroepidemiological studies of HSV-2 (see table 1). Further details of these studies are given elsewhere.^4,9,10

Developing a risk score for each study

We used standard univariate statistical tests to identify sociodemographic, behavioural, and clinical factors significantly associated with antibody to HSV-2. Forwards stepwise selection was then used to select the best of these factors to form a logistic regression model for HSV-2. The log odds ratios of the factors in the model are taken to indicate their importance, so that an additive scoring system results. The value of one unit of risk on the log odds scale was then selected so that as many factors as possible were close to a whole number of units and no factor represented more than five units. Each factor was then assigned a number of risk units, equal to the log odds ratio divided by the unit value, rounded to the nearest whole number. The overall risk score is then the sum of these terms.

For example, in study 4, one unit of risk was chosen to be 0.65 on the log odds scale. For number of lifetime sexual partners, relative to 0–2 the log odds ratios for 3–4, 5–9, and 10+ were 1.43, 2.10, and 2.80 respectively. Dividing by 0.65 and rounding, we see that 3–4, 5–9, or 10+ partners represent 2, 3, or 4 units of risk respectively.

The sample HSV prevalence is calculated for each value of the risk score. The score is then grouped to capture as much of the variation in prevalence as possible, subject to no more than five groups in total and each containing at least 8% of the sample to allow accurate estimation of the prevalence. However, for study 3, owing to similar prevalence across the middle risk scores, only four groups are created of which the highest risk group contains only 4% of the sample but has a substantially higher prevalence. For each risk group the sample prevalence and 95% confidence interval are determined.

Assessing similarity of associations with HSV-2 across settings

To consider whether a common scoring system could be used in different settings around the world, we assessed the similarity of associations with HSV-2 in the different studies. Age, sex, and number of lifetime partners were collected in each study and associated with HSV-2. Odds ratios for these factors, adjusting for the other two, were calculated using logistic regression. The similarity of the associations (that is, odds ratios) across studies can be formally assessed by testing for the main effects of study and interactions with the other factors.

All analysis was performed in spss 10 or stata 6.

RESULTS

Developing the risk scores

We include detailed results for study 1 in order to illustrate how the algorithm was constructed; results for other studies are available from the authors on request. Sex, age, genital blisters, lifetime number of partners, history of non-specific urethritis (NSU), gonococcus, warts, lice, Trichomonas vaginalis (TV), and thrush are all significantly associated with HSV-2. Country of birth, ethnicity, marital status, oral blisters/sores, cold sores, mouth ulcers, sexuality, age at first sex are not. Table 2 indicates the factors identified in the model selection process. The interactions between sex and the other three terms were each tested, and none found to significantly add to the model. The score to be added for each factor is specified in the final column of table 2. For example, if the patient is a 23 year old (0 points) man (0 points) with three lifetime partners (2 points) and no history of genital blisters (0 points) his score is 2, whereas if the patient is a 34 year old (2 points) woman (2 points) with seven lifetime partners (2 points) who has never had genital blisters (0 points) her score is 6.

The resulting scores, ranging from 0–10, were grouped according to the HSV-2 prevalence into five levels: 0–3, 4, 5, 6, and 7 or more. In table 3 the prevalence of antibody to HSV-2 is presented by risk group. In study 1, for example, the prevalence increases dramatically from 0.7% to 46.3% across the risk groups. The corresponding figures are also presented for the other studies.

Assessing similarity of associations with HSV-2 across settings

After controlling for the three risk factors, studies 2 and 4 have significantly higher HSV-2 prevalence than the other studies. Relative to study 1, the study odds ratios (95% CI) are 3.18 (2.38 to 4.26), 1.82 (1.39 to 2.37), and 3.20 (2.39 to 4.29) for studies 2, 3, and 4.

Furthermore we found the associations with sex (χ² = 15.6, 2 df, p = 0.0004), and number of partners (χ² =17.5, 6 df, p = 0.008), differed significantly across studies. Table 4 shows that, relative to study one, in study 2 the association between sex and HSV-2 is weaker, and in study 3 the association between number of partners and HSV-2 is weaker. The association with age did not vary significantly (χ² =8.6, 6 df, p = 0.20) across studies.

CONCLUSIONS AND FURTHER WORK

We have seen that it is possible to develop a risk score, which helps to predict an individual’s risk of HSV-2 on the basis of only a few pieces of information in all four study settings. These data will usually be gathered routinely during sexual history taking. Each of the resulting risk groups contains an appreciable proportion of people tested, yet the prevalence of HSV-2 varies sufficiently across groups in each study to have a substantial impact on the post-test probabilities of infection (following either a positive or negative test), and hence test interpretation. This is illustrated in table 5, where three realistic tests are considered.³ For example, someone with a low risk score should be greatly reassured by a negative test result as their post-test probability of infection is also low. If they have a positive result, however, they should be aware that this has a significant likelihood of being false, as their post-test probability will be substantially less than 100%, even using a test with 98% sensitivity and specificity. The converse is true for individuals at high risk of infection. In addition to informing test interpretation, the patient’s pretest probability can be used in advance of testing. For example, where it is clear that a patient’s post-test probability will be inconclusive (for example, 20%–80%) whether the test result is positive or negative, then the clinician and patient can discuss whether the test is likely to be of benefit or not. At present many physicians are not offering testing for HSV-2 antibody on the basis that they may do more harm than good (for example, by erroneously telling someone that they have a lifelong, viral, sexually transmitted infection). By using a simple algorithm such as the ones that are outlined here, preferably in conjunction with the patient, it is less likely that false results will be given. Additionally it may help the patient understand the limitations of testing in certain circumstances.

Alternatively, in those few laboratories able to offer confirmatory HSV-2 antibody testing using a western blot assay, this approach could be used to decide which samples to retest using western blot. (The HSV-2 antibody western blot assay is expensive and labour intensive and unlikely to become widely or commercially available.³)

We have not evaluated this approach for HSV-1 testing as we feel this is a less controversial area. The overall prevalence of HSV-1 infection in the community is higher so there is less likelihood of false positive and negative diagnoses. Additionally, the implications of either a false positive or negative diagnosis are likely to be less important. HSV-1 is generally considered a minor, viral illness and even if genitally acquired is unlikely to lead to recurrent genital herpes.

Since the broad themes of sexual behaviour, age, sex, and clinical symptoms are associated with HSV-2 prevalence in every setting it is tempting to feel that a common risk scoring can be developed and applied identically at different sites. This would certainly be of greatest practical use. It is true that if the risk score from one setting is applied to another then it is likely to describe a substantial amount of the variation in the second setting. For example, we applied the risk scoring from study 1 to individuals in study 2 and found that the prevalence in the risk groups increased from 1.4% to 50.7%.

However, our findings that three key behavioural and demographic risk factors have significantly different associations across studies, reflecting the different make up of the populations concerned, and that study itself is a risk factor, suggest that a common scoring system based on such factors is not possible. In general some individuals will be assigned to inappropriate risk groups—for example, those with a major risk factor in one setting which is a mild risk factor in another. In addition, it is not possible to say exactly what the prevalence will be in the risk groups when a scoring based on one setting is applied to another. This finding is perhaps not surprising, firstly, because it is impossible to measure the exact behavioural risk experienced by an individual since it requires information unlikely to be available to the patient—for example, detailed risk information for each lifetime partner. Secondly, a related problem is that the association between sexual behaviour and HSV-2 prevalence is dependent on the prevalence in the population, which will vary across settings.

This work to assess whether a common risk scoring can be used across settings is limited by a lack of common patient information across studies, few clinical data collected, and the small number of studies included. In particular, a greater degree of commonality might be expected if the scoring includes a variety of clinical factors. Further projects might be designed without these limitations.

Further work might also address whether more complex approaches to selecting risk factors and combining this information into a scoring system would provide additional benefit. More complex approaches would be feasible, in particular, if the calculation of pretest and post-test probabilities is computerised.

Importantly, further work is also needed to establish whether using such an evidence based approach to HSV-2 testing such as the one described would actually benefit patients and clinicians in practice.