Background
It is common knowledge that individuals in any population vary in their likelihood of contracting acute infectious disease. Inherited patterns of innate immunity responses [
1,
2], variable adaptive immunity due to previous exposures to causative agents of the infections [
3], and modification of the above responses by chronic diseases or their immune suppressive treatment [
4] are each plausible explanations for a part of the observed differences in individual susceptibility. However, apart from a possible chronic disease record, this background information is not usually available for identifying suitable persons to participate in an infectious disease trial, or for trial-arm matching. In order to improve the harmonization, one can try to take advantage of expected post-randomization events affecting the emergence of infections including predicted frequency and intensity of exposures to infectious agents. Factors such as being a parent of young children regularly visiting a day care center, and occupational exposure to persons suffering from active infectious disease are often considered as risk factors for infectious disease in adults [
5]. Yet, no generalizable data are available for relative roles or quantitative assessment of the different designated risk factors.
In this paper we describe an attempt to develop a prognostic model for predicting individual incidences of acute respiratory tract infection (RTI) or gastrointestinal tract infection (GTI) in the general adult population. The model is based on data collected in a controlled, cluster-randomized intervention trial evaluating the efficacy of intensified hand hygiene on occurrence of RTI or GTI in adult office workers in 2009–2010 in the Helsinki region, Finland (The STOPFLU Study). Occurrence of symptoms of RTI or GTI and that of recognized exposures to other persons with symptoms of RTI or GTI, were collected by weekly electronic reports. The protocol and the main results of the trial have been published before [
6,
7], and selected parts are described in Additional file
1. In order to be able to homogenize the study arms in the original trial, interview data for various designated risk factors for acute infection were collected from the volunteer participants. In the current study, this questionnaire data was subjected to multivariate analysis in order to identify potential predicting factors, and significantly contributing variables were used in creating a tentative prognostic model. Model-predicted individual incidences of both infections and exposures were then calculated and compared with the observed ones.
Discussion
In this paper we describe an attempt to develop prognostic models for predicting future incidences of acute respiratory or gastrointestinal tract infections in the general adult population. Both the potential predicting factors and the endpoint data used were derived from a previously published [
7] hand-hygiene intervention trial carried out in Finland in 2009–2010. The potential predicting variables tested here were based on a prerandomization questionnaire used in the trial in order to harmonize the trial arms. Age, gender, chronic cardiovascular or respiratory disease, regular use of public transport, and history of seasonal influenza vaccination significantly influenced the subsequent GTI and/or RTI incidences. However, prognostic models based on these factors had only a moderate predicting power as regards the RTI incidence and a poor predicting power in the case of GTI incidence.
Caveats relating to data based on self-reporting only, and not confirmed in individual cases by observations or examinations by trained health care personnel, have been discussed before [
7]. A questionnaire enquiring into a range of personal matters in the context of an occupational health trial may also have been felt too intimate and, in some persons, result in intentional or unintentional adjustment of the answers. The emergence of a symptomatic acute infection in the respiratory or gastrointestinal tract requires a transmission of an infectious agent, usually from another person, but a relevant exposure can be unrecognizable, due to an asymptomatic infection in the contact person [
14], or can for some reason remain unnoticed, ignored, or forgotten by the time of the weekly reporting.
Another weakness of the current study is its post hoc nature which results, among other things, in the fact that some of the generated variable-specific subgroups are likely to be too small to provide sufficient power to detect all possibly relevant influences. A further point to note is a possible selection bias. Only about half of the eligible employees in the trial clusters volunteered to participate in the study. Reasons for declining the invitation to the study were not examined but one can assume that there may be differences in the health behavior and other personal properties between these “silent” and the active, reporting members of the trial clusters. As a consequence, we cannot be sure how well the results of the current study represent the general adult population.
A common practice in analyzing clinical trials is to present the results in both “per-protocol” and “intention-to-treat” modes with the latter often also including both dropouts and cases with missing supplementary data. The STOPFLU Study protocol included a plan to include in the analysis those with intermittently lacking weekly reports (usually due to holidays, but in practice, for any reason), participants recruited after the onset of the intervention (“missing” early data), as well as reports from participants who ceased to report before the end of the trial (“dropouts”). The latter was expected to occur relatively frequently because the employing corporations of the participants were in the middle of reorganization (Additional file
1). In spite of their per-protocol nature these defects in the data require comments for potential influence on the developed models. Missing data in the potential predicting variable questionnaire were minimal. Usually only one or a few persons out of 717 had declined answering a given question. An exception was “passive smoking” which was not answered by nine persons (Additional file
2). It is highly unlikely that these omissions could influence the outcome, and thus they were ignored in the model building. Intermittent breaks in weekly reporting were relatively common but assumed to take place at random and, therefore, also excluded from the model variables. Missing data, due to delayed enrollment or premature ceasing of reporting, concerned about 30% of the participants. The dates of the delayed enrollment were distributed over the duration of the trial (Additional file
3) and are, therefore, unlikely to influence the outcomes significantly. Ceasing of reporting was almost always related to stopping working for the previous employer or due to transfer to another working unit. Therefore, we initially considered that it was not endpoint dependent and thus not affecting the accuracy of the outcomes. According to a regression analysis [
15], ceasing the weekly reporting was not dependent on reported endpoint incidences, and thus, from a statistical point of view, occurring at random. The entire duration of the trial was 16 months but for some individuals the total reporting time was relatively short (Additional file
3). Since in theory, short reporting periods might influence the endpoints if occurring at a low or very high epidemic season, we repeated the analysis for identifying significant predictor variables by excluding participants with fewer than 50 weekly reports. The resulting variable-specific IRRs did not much differ from those of all reported data although the
p values in the case of GTI increased beyond the significance limit (see Additional file
3).
The overall predicting power of the models developed was at best only moderate for RTI and poor for GTI. This was not surprising after seeing the IRRs of the identified significant predicting variables which differed relatively little from 1 and had broad confidence intervals. The tested variables were no better in predicting the incidence of exposures to other persons with RTI and only vaguely for exposures to GTI. Temporally associated exposures were previously reported to have a strong relation to homologous disease in the reporter [
14] but the temporal association cannot be predicted in studies of long duration. Out of the different factors included in the analysis, only a few showed a statistically significant contribution to the endpoint. Women had reported relatively more exposures both to GTI and RTI as well as weeks with GTI symptoms than men. We believe that this difference was most likely based on plausibly different behavior of the genders rather than caused by a true gender-related biological factor. It is possible that women have had more contacts with other people during their free time than men. Women may also be relatively more sensitive to recognize and report possible exposures. As regards GTI, the observation might also reflect real life as it is possible that women, as mothers, are often in more close contact with their sick children, a likely source of GTI in many cases. Through most of the age ranges of the participants, increasing age appeared to reduce the risk of exposure to both GTI and RTI as well as that of the emergence of RTI and GTI symptoms. The observed influence of age on the incidences of both exposure and infection is, as such, understandable. Younger people move around more than the older ones during their free time and have more contacts with other people. It is also possible that younger people are relatively more sensitive to recognize and report possible exposures and to notify symptoms of infection. Of course, accumulating acquired immunity to some of the infectious agents could contribute to this decreasing trend of symptomatic RTI and GTI.
Three other variables were associated with an increased RTI incidence. Various surfaces in the public transport vehicles may serve as invisibly contaminated fomites, and infections may be transmitted manually without recognition of the exposure [
16,
17]. As short distances to infected other persons in the vehicles are difficult to avoid and, as respiratory viruses may also be spread via aerosols without coughing [
16‐
18], it is understandable that using the transport was found to increase the risk of RTI. Chronic cardiovascular or respiratory disease is a known risk factor for complications of influenza and most likely for other respiratory viral infections [
19]. Hence, the almost significantly increased incidence of weeks with reported RTI symptoms but without an increase in the reported exposures is not unexpected. Influenza vaccine recipients during the preceding season partially overlapped the above subgroup suffering from chronic disease (these people get the influenza vaccine cost-free in Finland). Possible specific reasons for taking the vaccine were not enquired, but influenza vaccine is recommended by the health authorities to people who want to get a protection against epidemic influenza. It is possible that the remaining vaccine recipients have historically suffered from frequent colds and have, therefore, obtained the vaccine. Anyhow, as above, a slightly increased incidence of RTI is not surprising. A protection effect through the trial was not expected, as virologically documented influenza was a minor element among the tested, relatively severe RTI during the winter 2009 epidemic [
7], and a vaccine received in autumn 2008 had no effect against the H1N1 virus pandemic during the following influenza season.
We were somewhat surprised to see that the statistical multivariate analysis revealed a lack of influence for two designated cluster-risk sum calculations: young children in the household and smoking. It is common knowledge that children suffer from acute RTI and GTI more frequently than adults and might easily transmit the infection to their parents who are often living in close physical contact with sick children. In the univariate analysis, having under-school-age children in the household or a child in outside-home day care were both associated with IRRs significantly above 1. The stronger effect by the age and gender may have diluted out these effects in the multivariate analysis. It is possible that a larger proportion of households with children in these groups could also have rendered the difference significant in the multivariate model. One hundred and eighty-five participants (about one quarter) lived in a household with young children, and in 95 out of 716 households (13.3%) at least one of the children was under school age.
Smoking was another variable where we expected, but did not obtain, an effect on the incidence of respiratory symptoms. It is well-established that smokers suffer from both acute and chronic respiratory symptoms more than nonsmokers of the same age [
20,
21]. The reasons for our failure to see the expected effects can only be speculated. According to the questionnaire answers, 150 out of 714 participants (21%) were smokers, a number well-corresponding to the known proportion of smokers in the general population in Finland [
22]. However, only 18 of the participants of the study can be considered as heavy smokers (more than 20 cigarettes per day).
Living with an adult nonparticipant, who might carry home infections from their job, and passive smoking were associated with increased reported exposure incidences but not with those of RTI or GTI symptom weeks. One can speculate that a recognized exposure has resulted in intensified hand hygiene and that the person has succeeded in avoiding the consequent symptomatic infection. Another possible explanation is a situation-prompted increased sensitivity to recognize and/or register the exposures. On the other hand, passive smoking is an irritating situation to most people who are not smoking themselves, and may result in a sort of over-sensitivity to detecting symptoms of disease in smokers.
Acknowledgements
The STOPFLU Study Group: Tapani Hovi, Carita Savolainen-Kopra, Thedi Ziegler, Terttu Korpela, Ali Amiryousefi, Pirjo Anttila, Jaason Haapakoski, Pentti Huovinen, Markku Huvinen, Heikki Noronen, Pia Riikkala, Merja Roivainen, Petri Ruutu, Juha Teirilä, and Erkki Vartiainen. The authors are grateful to Ali Amiryousefi for preliminary statistical analysis of this dataset.