Our study suggests that signs and symptoms of disease do not predict acute human Q fever in GP patients, confirming earlier results from a Dutch study in hospitalized patients.(11) However, signs and symptoms may be useful in ruling out acute Q fever in favour of other acute lower respiratory tract infections. This is especially relevant in cases where pneumonia is not suspected and a non-prescribing or delayed prescribing approach would seem more appropriate, helping reduce the overuse of antibiotics. In the cohort of patients tested in our region, this would have been particularly relevant in the immediate post-outbreak phase where the number of tests for acute human Q fever remained high but the proportion of seropositive cases was very low (7%), and prevalence of pneumonia was also low (12%). Even during the outbreak phase, only 26.2% of tested individuals were positive, and ruling out acute Q fever by symptoms would likely have reduced overuse of antibiotics.
Male sex, fever, and pneumonia were positive predictors of acute Q fever across all four of our models, in accordance with other studies.(2, 20) Cough was a negative predictor in three models, suggesting that cough as a symptom may be particularly useful in ruling out Q fever in suspected cases. Cough is considered a common symptom in upper respiratory tract infections. Its presence, according to our findings, may point to respiratory conditions other than Q fever.(21) Specifically, cough has been described as a symptom suggestive of influenza, rather than, for example, common cold.(7) Overall, in our sample cough was the most prevalent symptom – second only to flu-like illness – in questionnaire respondents from the second wave, both in Q fever positive and Q fever negative subjects. This, combined with the low rate of Q fever positive findings during the second wave, may suggest that the rise in Q fever testing activity by GP’s and specialty physicians during the second half of 2009 and the early months of 2010 may – at least to some degree – have resulted from patients presenting with respiratory symptoms due to increasing pandemic influenza A(H1N1)pdm09 activity in that period. Moreover, due to long persistence of anti-Coxiella phase II IgM following infection, some of the subjects who tested positive during the second wave may have been misclassified as acute Q fever. While abdominal pain was a negative predictor of acute Q fever across all four models, gastrointestinal symptoms such as abdominal pain and diarrhoea were much less prevalent than cough in both Q fever positive and Q fever negative subjects from both waves, and the nature of the observed negative association of abdominal pain with Q fever remains unclear. Studies on the gastrointestinal symptoms in patients with influenza report prevalence rates ranging from 0.6 to 6.6% for influenza A(H1N1) infections, and 9.8 to 57.5% for influenza A(H1N1)pdm09 infections, suggesting a possible association of gastrointestinal symptoms in our study with the 2009 swine flu.(22)
Use of signs and symptoms of disease to rule out acute Q fever would be most appropriate in patients with non-severe lower respiratory tract infections, i.e., in cases where pneumonia is not suspected clinically. In such cases, use of antibiotics has been shown to provide little benefit in primary care, both overall and in patients aged 60 years and above, but may cause slight harms.(23, 24) In cases with clinical suspicion of pneumonia, however, the benefit of antibiotics would outweigh potential harms. Whereas doxycycline is recommended only as a second line treatment for empiric management of community-acquired pneumonia (CAP) in several national guidelines, it is a first-line antibiotic in others, and generally considered to be safe and effective. Thus, in cases of lower respiratory tract infections where acute Q fever is included in the differential diagnosis and pneumonia is suspected, use of doxycycline would seem a perfectly justifiable choice.
Predictive values are greatly impacted by prevalence of the disease in the base population. Positive predictive values (PPV) tend to be low in situations where prevalence in the base population is low, as was the case in our study, where post-outbreak seroprevalence of prior exposure to C. burnetii in the base population was estimated a mere 2.9%.(1) With PPV ranging between 4.6% and 8.3%, mirrored by low areas under the Receiver Operator Curves, our models had no use as a diagnostic tool for acute Q fever. Conversely, negative predictive values (NPV) tend to be high under circumstances of low disease prevalence. With NPV ranging between 91.7–99.5%, our models were able to rule out the presence of acute Q fever with a relatively high degree of confidence. Nevertheless, decisions favouring a delayed or non-prescribing approach should ideally be corroborated by other clinical findings supporting such approach. In other contexts, for example in a well-circumscribed population of patients with high-risk exposure to a known source, prevalence may be (much) higher, with resulting decline in NPV.
To the best of our knowledge, ours is the first study to use post-outbreak data to validate prediction models for acute Q fever derived from outbreak data, thus enhancing the generalisability of our findings. Moreover, our study is first to assess the predictive potential of signs and symptoms for the diagnosis of acute Q fever in a large population of subjects most of whom were GP patients. Other studies attempting to predict Q fever by signs and symptoms, including a retrospective case-control study from the Netherlands, were performed in hospital settings. The Dutch study reported that clinical signs and symptoms were not helpful in differentiating adult hospital-referred patients with acute Q fever from a hospital-referred control group.(11) A second study aimed to predict Q fever in patients presenting with community-acquired pneumonia to the hospital. The only symptom independently associated with Q fever in this study was headache. The prognostic score derived from multivariate logistic regression included male sex, age 30–60 years, a low leukocyte count and a high C-reactive protein (CRP) level, along with headache, as predictors of Q fever pneumonia.(10) A third study attempted to predict acute Q fever in febrile patients from rural Kenya, based on parameters including a range of clinical signs and symptoms. The study identified acute lower respiratory infection, abdominal pain, diarrhoea and a history of fever lasting > 14 days as independent significant positive predictors of acute Q fever. A prediction score derived from a modelling approach similar to ours was reported to reliably identify acute Q fever in febrile patients with undifferentiated illness.(9)
Our study had a number of limitations. Selection of subjects for inclusion in our study was based on laboratory Q-fever testing outcomes rather than random sampling, with a potential for selection bias, e.g., due to variations in diagnostic strategies between individual physicians. Laboratory confirmed cases of acute Q fever and patients who were Q fever negative were both selected based on signs and symptoms leading to addition of Q fever in the differential diagnosis, possibly resulting in some weakening of the association under study. Our laboratory data were strictly limited to outcomes of Q fever testing, precluding us from assessing signs and symptoms in relation to possible alternative outcomes. As mentioned above, misclassification of positive laboratory results as acute Q fever infection cannot be entirely ruled out, since phase-II IgM antibodies to C. burnetii, which at the time of the outbreak were considered to be reliable markers of acute Q fever infection, have been shown to persist for longer periods in individual patients, thus complicating the differentiation between past Q fever infections and acute respiratory infections with different etiologies.(25) Cross-validation and testing of our models were performed on samples from the same base population, potentially compromising generalizability of our findings. Lack of external validation of our models, however, may have partly been offset by the fact that we performed validation against a holdout sample, i.e., data from the second wave of Q fever testing. Testing during the second wave took place in what may be described as a immediate post-outbreak transition period where Q fever was increasingly replaced by other aetiologies of clinical respiratory disease, thus distinguishing the population of individuals tested during the second wave from those included in the first wave. Splitting our first-wave dataset for internal cross-validation may have resulted in loss of power, and may have contributed to discrepancies between our four models in terms of predictors included in each model. Nevertheless, all four models performed equally well in terms of their negative predictive value, suggesting they may be equally useful as predictive tools. Whereas the Youden index is a commonly used method for cut-point selection in ROC analysis, there a several other approaches, whose application may have led to different results.(26)