Causes of fever in primary care in Southeast Asia and the performance of C-reactive protein in discriminating bacterial from viral pathogens

Highlights • Evidence on causes of fever is limited in Southeast Asia, especially in primary care.• Point-of-care diagnostic tools could guide health workers’ antibiotic prescription.• C-reactive protein was significantly increased in cases of bacterial infections.• Most primary care patients recovered regardless of antibiotic prescription.• Antibiotic prescription should be an exception in the primary levels of care.


Introduction
Fever is a common reason for seeking health care in Southeast Asia. As malaria incidence declines, bacteria and viruses now represent the main contributors to acute febrile illness (World Health Organization, 2017;Chheng et al., 2013;Mayxay et al., 2015;Southeast Asia Infectious Disease Clinical Research Network, 2017;Wangdi et al., 2019). Identifying these pathogens is challenging, even in well-resourced laboratories with specialised staff, and most aetiological data for febrile illness originate in tertiary hospitals (Shrestha et al., 2018). Hospitalised patients are by definition more severely ill, often with important comorbidities, implying that findings may not be applicable to febrile patients attending primary levels of care.
Primary care in low-middle income countries is typically characterised by a shortage in human resources, diagnostics and evidence-based guidelines (Mendelson et al., 2016). Studies investigating causes of fever in this environment are few and frequently of poor quality: enrolment is often limited to a single clinical presentation and specific age category, and microbiological investigations rarely use gold-standard methods (Mueller et al., 2014;Kasper et al., 2012;Capeding et al., 2013). Additionally, most patients attend primary care early after symptom onset with nonsevere presentations, lowering the chances of detecting a pathogen (Mueller et al., 2014;Kasper et al., 2012;Capeding et al., 2013;Schmidt et al., 2014;Cunnington, 2015;Waterer and Rello, 2011;Hackett et al., 2002;D'Acremont et al., 2014;Hildenwall et al., 2016). Empiric treatment guidelines are therefore based on limited epidemiological evidence and are often implemented by insufficient and poorly trained staff, contributing to irrational antibiotic prescription practices (WHO, 2006;Chen et al., 2004;Kanchanachitra et al., 2011).
High prescription levels are partly driven by frequent clinical overlap between bacterial and viral infections, challenging the identification of patients who might benefit from antibiotics (Chheng et al., 2013;Mayxay et al., 2015;Koh et al., 2010). Given these limitations in clinical judgment and laboratory structures, point-of-care testing (POCT) to guide fever management could be beneficial in primary care settings (WHO, 2011). Pathogen-specific tests represent one such option but several barriers undermine their potential use: these only exist for a small number of pathogens, with inconsistent performance, and most are antibody detection-based that might preclude the distinction between active infection and past exposure (Saraswati et al., 2018;Picardeau et al., 2014). A few antigen detection-based POCTs exist but their integration in low-level care is unrealistic: a Salmonella typhi rapid test requires laboratory infrastructure with poor detection in blood, even at high concentrations, and results are not available before 24-48 h (Castonguay-Vanier et al., 2013;Kuijpers et al., 2018); test sensitivities for influenza virus A, respiratory syncytial virus (RSV) and group A Streptococcus antigen-based POCTs are inconsistent (Drexler et al., 2009;Trombetta et al., 2018;Chartrand et al., 2015;Cohen et al., 2016); and accurate dengue antigen-based RDTs have not been found to be cost-effective in resource-poor settings (Lubell et al., 2016;Lim et al., 2017).
Non-specific host biomarkers measure the host-response to stimuli, and have been evaluated in the context of fever to discriminate between bacterial and viral pathogens (Kapasi et al., 2016). C-reactive protein (CRP) is one of the most studied hostresponse biomarkers of bacterial infection, consistently showing high sensitivity and moderate specificity, and CRP POCTs have been shown to be cost-effective in resource-poor environments (Lubell et al., 2016;Kapasi et al., 2016). However, 80% of studies evaluating CRP performance originate from high-income countries (Kapasi et al., 2016). In Southeast Asia, these evaluations are mainly hospital-based (Sutinen et al., 1998;Choo et al., 2001;Wangrangsimakul et al., 2018), with limited evidence at the community level, community-based study (Lubell et al., 2015). Good diagnostic performance of CRP in identifying bacterial infections was observed, but generalisability was limited due to demographic, clinical and diagnostic heterogeneity of these studies.
This study aimed to identify key organisms among acutely febrile children and adults attending primary health care in Southeast Asia, and to evaluate the performance of CRP for discriminating between bacteria and viruses.

Study sites
Chiang Rai province is the northernmost province in Thailand and borders Myanmar and Lao People's Democratic Republic. The majority of the population are Thai, with approximately 15% ethnic minorities and hill tribes. The six participating primary care sites were located within a 30-km radius of Chiang Rai city centre, covering rural and peri-urban as well as mountainous and plateau areas.
Hlaing Tha Yar, Lower Myanmar, is a peri-urban township on the west side of Yangon. The township has the highest rates of diseases related to hygiene and environmental conditions (e.g. diarrhoea, dysentery and tuberculosis) in Yangon (Htwe et al., Figure 1. Study flowchart. RNP3, Ribonuclease P3 gene as an extraction and specimen integrity control. 2017). Four sites were included: three primary care clinics and one outpatient department from a public governmental hospital.
Both Chiang Rai and Hlaing Tha Yar are defined by a tropical climate.

Study design
Specimens were collected from febrile patients recruited into a previously described multi-centre randomised controlled trial evaluating the impact of CRP testing on antibiotic prescription in primary care (Althaus et al., 2019). Febrile children and adults (defined as 12 years of age) were recruited between June 2016 and August 2017. Inclusion criteria were being aged 1 year with a documented fever (defined as a tympanic temperature >37.5 C) or a chief complaint of acute fever (<14 days), regardless of previous antibiotic intake and co-morbidities other than malignancies. Exclusion criteria were symptoms requiring hospital referral, defined as: impaired consciousness; an inability to take oral medication or convulsions; a positive malaria test; the main complaint being trauma and/or injury; a suspicion of either tuberculosis, urinary tract infection, local skin infection or dental abscess; any symptom present for >14 days; any bleeding; an inability to comply with the follow-up visit at day 5.
On the day of enrolment, all patients had demographic information collected and underwent a routine clinical examination, including vital signs such as blood pressure, pulse, respiratory rate, and temperature. Patients were followed-up after their enrolment both at day 5 and day 14.

Specimen management
Of the 2392 febrile children and adults recruited, 799 were randomly allocated to the control group with blood specimens collected for off-site CRP testing (as compared with the intervention groups that had CRP tests performed on-site). Details are illustrated in Figure 1. Antibiotics were prescribed to this group according to routine clinical practice; clinicians were not informed of the CRP results or any aetiological findings.
Microbiological investigations were carried out at Mahidol-Oxford Tropical Medicine Research Unit (MORU) in Bangkok, Thailand, and in the Amsterdam Public Health Laboratory, the Netherlands. A combination of molecular and serological assays were used to detect 35 bacteria, 30 viruses and three fungi from plasma and nasopharyngeal (NP) swab specimens (Supplementary figures and Tables 1 and 2).
Among the 740 patients for whom a blood specimen was collected, a sufficient volume of plasma was available to test 656 of them using the Rickettsial immunofluorescence assay (IFA); 626 using the bacterial singleplex polymerase chain reaction (PCR) including 16S, Rickettsia spp., Orientia tsutsugamushi and Leptospira spp.; and 601 using the Taqman array card (TAC) assay. The remaining plasma and the DBS collected among Myanmar children aged <5 years was used for the viral singleplex PCR on 678 patients, screening for dengue, chikungunya and zika viruses (Supplementary Table 3). CRP concentrations (mg/L) were measured on the same day of sampling using a quantitative CRP reader (NycoCard II Reader, Axis Shield, Norway), with a lower limit of detection of 8 mg/L. Comorbidities included HIV, chronic hepatitis B or C, cirrhosis, diabetes mellitus, asthma, anaemia, chronic obstructive pulmonary disease, gastritis, congenital heart or kidney disease, alcoholism, dyslipidaemia, G6PD deficiency, hypertension, rheumatic heart disease, ischaemic heart disease, thalassaemia, thyroid disease, and Alzheimer's disease. Neurological symptoms included headache, confusion or hearing loss. Respiratory tract symptoms included sore throat, dyspnoea, pain on inspiration, runny nose, or cough. URT symptoms (upper respiratory tract symptoms) were defined by the presence of either runny nose, sore throat or cough. Gastrointestinal symptoms included nausea, vomiting, jaundice, diarrhoea, or abdominal pain. Other symptoms were defined by the presence of fever alone or symptoms that were not neurological, respiratory or gastrointestinal. Common symptoms in this group included myalgia, arthralgia, tiredness, chills, sweating, weight loss, skin eruption, dysuria, dizziness or eye redness.

Attribution of causality
A conservative diagnostic approach was chosen, following the strength of evidence for each result: the bacteria and viruses detected in blood with the molecular and serological methods using the thresholds detailed in Table 1 were considered "target" organisms. Considering NP swabs, only influenza virus (A and B), respiratory syncytial virus (RSV), human metapneumovirus, and Bordetella pertussis were included as target organisms, given the higher likelihood of causality, according to available evidence in the literature (Supplementary Table 4). Other organisms detected in NP swabs were not classified as causal, because pathogenicity is not consistently demonstrated in children or adults. In case of codetection in blood and NP swabs, target organisms detected in blood were assumed to be the primary cause of illness.
Patients were allocated to the bacterial or viral group if a bacterium or a virus was detected in blood. Considering NP swabs, only bacteria and viruses considered target were allocated to the bacterial or viral group.

Statistical analysis
Organism distribution among children and adults (defined as 12 years of age) were described separately. Descriptive analysis for continuous variables with normal distribution used means and standard deviations (SD) and medians with inter-quartile ranges (IQR) for non-normally distributed continuous variables. Comparison between groups used t-tests for normally distributed variables, the Mann-Whitney test for non-normally distributed variables, and Chi-squared test for categorical variables. Table 2 Bacterial and viral detection in blood and nasopharyngeal swabs specimens among children and adults in Chiang Rai, northern Thailand and Hlaing Tha Yar, Lower Myanmar, 2016-2017. Organisms with a high likelihood of causality in nasopharyngeal swabs are in bold.
CRP values were compared across aetiological groups and clinical syndromes using the Mann-Whitney U test for two-group comparisons and the Kruskal-Wallis test for multi-group comparisons. Non-parametric receiver operating characteristic (ROC) curves were plotted and the Wald test was used to compare areas under the curve. Covariates included patient age and prior use of antibiotics. Diagnostic accuracy was assessed by calculating the areas under the ROC curves (AUC). An AUC of >0.9 was considered excellent; 0.8-0.9, very good; 0.7-0.8, good; 0.6-0.7, average; <0.6, poor (Ying et al., 2011;Campbell, 1994). Sensitivity, specificity and percentage of correctly classified cases were also assessed for the two CRP cut-off points used in the original trial: 20 mg/L and 40 mg/L; these were compared with the accuracy of routine prescribing practice (Althaus et al., 2019). Data analyses were performed with STATA version 15 (College Station, Texas, USA).
Ethical approval and role of the funding source

Baseline characteristics
Of the 799 patients prospectively enrolled and randomised into the trial control group, 773 (96.8%) had at least one blood or NP swab specimen collected, including 371 (48.0%) children and 402 adults (52.0%). Of these 773 patients, 33 had only a NP swab and no blood collected, while 146 had only a blood specimen collected without an NP swab; among these, 80 had a blood specimen obtained on a dried blood spot (DBS). As shown in Table 1, children presented significantly earlier after symptom onset than adults, with fewer comorbidities and less self-reported medication (p < 0.05). Antibiotic intake declaration was similar in children and adults (p = 0.237).
Clinically, respiratory symptoms were the most prevalent both in children and adults. Within patients with respiratory symptoms, the most frequent were localised in the upper respiratory tract, with common cold being diagnosed in 48.2% (120/249) of children and 45.0% (108/240) of adults (p = 0.479). Gastrointestinal symptoms were the second most prevalent presentation, and there were no differences between children and adults.
Elevated CRP defined as 50 mg/L in children and 100 mg/L in adults. SAE: serious adverse event, defined as admission to hospital or death within 14 days of enrolment. Broad-spectrum antibiotics included ceftriaxone, cefixime, ciprofloxacin, levofloxacin, azithromycin, and amoxicillin with clavulanic acid.

Patient outcomes by aetiological group
No evidence for a difference in antibiotic prescription was observed between the bacterial and viral groups at day 0, and clinical outcomes were also not significantly different between the two groups (Table 3).
Among patients with a bacterial organism, two-thirds did not receive any antibiotic (24/36, 66.7%), and those antibiotics received were often ineffective for the detected bacteria (Supplementary Table 6). Of the 12 patients with either Leptospira spp. or Rickettsia genus, nine (75%) did not receive any antibiotic. No evidence for a difference in clinical outcomes was observed after 14 days of follow-up, regardless of whether an antibiotic was prescribed.

Accuracy of current prescribing practices and CRP-guided treatment
Of the 227 patients with a target organism detected, 36 (15.9%) were allocated to the bacterial group and 191 (84.1%) to the viral group. The median CRP concentration was higher in the bacterial group compared with the viral one (18 [10-49] mg/L versus 10 [8-22] mg/L, p = 0.003, respectively). Among the 422 patients for whom no organism was detected or the organism was not considered causal, the median CRP was 10 (8-29) mg/L lower than among patients with a target bacteria (p = 0.004) but similar to patients with a target viral organism (p = 0.826). The distribution of CRP concentrations is presented by aetiological group in Figure 2.
The AUC for CRP distinguishing bacterial from viral target organisms was 0.652, 95% confidence interval (0.553-0.750), classified as average, as illustrated in Figure 3.
In the absence of CRP-testing, the sensitivity of current prescribing practices for detecting bacterial target organisms was 33.3% (95% CI 18.6-51.0%) and specificity was 60.2% (95% CI 52.9-67.2%). Using a threshold of 20 mg/L would imply an increase in sensitivity to 47.2% (95% CI 30.4-64.5%) and an increase in specificity to 68.9% (62.3-75.5%). Use of the higher threshold of 40 mg/L would imply a lower sensitivity of 27.8% (14.2-45.2%) and a higher specificity of 88.7% (83.3-92.6%). In total, based on current prescribing practices during the study, 57.8% of patients were correctly classified for the prescription of an antibiotic, while CRPguided treatment would have increased this to 65.8% using a 20 mg/L threshold and to 79.3% using a 40 mg/L threshold.

Discussion
This study investigated the spectrum of organisms among febrile children and adults in the community and evaluated the performance of CRP in distinguishing bacteria from viruses, including its potential impact on antibiotic prescription compared with current practice. Patients were prospectively recruited across ten sites in Thailand and Myanmar, including urban, semi-urban and rural areas, spanning over a full calendar year.
Leptospira spp., influenza virus, RSV and dengue virus were the leading organisms identified, which is consistent with previous reports in the region (Mueller et al., 2014;Lubell et al., 2015). The broad inclusion criteria, allowing for enrolment of all patients aged >1 year regardless of previous antibiotic intake, comorbidities (apart from malignancy), or clinical presentation, made the findings more generalisable than previous studies.
Investigating non-malarial acute febrile illness remains challenging in resource-poor areas (Mueller et al., 2014), and despite screening for multiple organisms on blood and respiratory specimens, it was only able to identify a probable cause of fever in 227 (29.4%) of patients. This low detection may be explained by the inclusion of only non-severe outpatients (Hackett et al., 2002;Jiang et al., 2017;Reimer et al., 1997), while other studies in Southeast Asia recruiting more severe and hospitalised patients identified an organism in around 50% of cases (Chheng et al., 2013;Mayxay et al., 2015;Southeast Asia Infectious Disease Clinical Research Network, 2017;Wangrangsimakul et al., 2018). Of the detected organisms, 15.9% (36/227) were bacteria, which may be explained by the lower risk of bacterial infections in non-severely ill patients, and where present, characterised by lower bacterial loads (Reimer et al., 1997).
Most bacteria were identified using a singleplex PCR and not the TAC assay, while viruses were equally detected by these two molecular methods. This lower sensitivity in the TAC assay for the detection of bacteria has been described in previous studies using multi-pathogen molecular detection platforms (Warhurst et al., 2015;Weller et al., 2012). The trade-off between advantages for screening multiple organisms at the same time with a simplified molecular platform should be weighed against potentially lower sensitivity, especially for bacteria such as O. tsutsugamushi or Leptospira spp., which are considered important drivers of acute febrile illness in Southeast Asia (Xu et al., 2017;Laras et al., 2002).
Conversely, the use of multiplex assays runs the risk of detecting organisms for which the causality and interpretation can be challenging: this study detected Klebsiella pneumoniae in the blood of eight patients, of whom six did not receive any antibiotic. It is unknown whether these patients had a transient self-limiting infection, carriage, or were false positive. A 2018 community-based study also used the TAC assay to identify infections among neonates in South Asia, but detected the presence of certain organisms among both controls and cases (Saha et al., 2018). A 2019 multi-country study into causes of severe pneumonia also excluded molecular assay results positive for Klebsiella pneumoniae because of poor assay specificity (O'Brien et al., 2019).
Furthermore, most of the patients presented with low CRP regardless of whether a bacterial or viral organism was detected, and recovered regardless of whether an antibiotic was prescribed. It is likely that invasive bacterial infections requiring an antibiotic are exceptions, while most primary care patients present with a self-limiting infection (Van den Bruel et al., 2007). Other primary care-based studies have recently supported restriction of antibiotic prescription to a small minority of patients: in Tanzania, a clinical trial using a 80 mg/L threshold lowered antibiotic prescription to 2.3% without affecting outcomes, while a study from the USA concluded that 72% of outpatients attending a general practice for a respiratory presentation should not even require a medical consultation, let alone an antibiotic prescription (Keitel et al., 2019;Renati and Linder, 2016). Strategies whereby testing for CRP as a predictor of clinical outcome rather than determining aetiology have been evaluated in primary care: a clusterrandomised controlled trial in Belgium showed CRP to rule-out serious infection using a 5 mg/L threshold, while a systematic review found CRP testing to be useful in identifying serious infections among febrile children (Verbakel et al., 2016;Van den Bruel et al., 2011).
In the current study, CRP performance in distinguishing bacteria from viruses was average (AUC 0.65) and lower than another study from the same region, which found AUC 0.83 among 1372 patients with a microbiologically-confirmed diagnosis from Thailand, Cambodia and Lao PDR (Lubell et al., 2015). It is likely that most bacteria detected in the current study caused a self-limiting infection, while previous CRP evaluations included more severe and hospitalised patients with bacteraemia detected by blood culture. Despite average performance, CRP-guided treatment would still result in more accurate classification of fever as bacterial or viral aetiology compared with current practices (assuming that prescription of antibiotics was indicative of health workers believing the illness to be of bacterial origin). This supports the findings from the original trial, where CRP testing safely reduced antibiotic prescription, which are in line with similar primary care-based studies both in Asia and Africa (Keitel et al., 2019;Keitel et al., 2017;Do et al., 2016).
This study had several limitations mostly relating to the limited scope and accuracy of the reference diagnostic tests, and the impact of even slightly less than perfect "gold-standard" reference tests on the evaluation of new diagnostic and biomarker tests can be profound (McHugh et al., 2019;Waikar et al., 2013). As mentioned above, the sensitivity and specificity of the multiplex TAC assay was not optimal for bacteria detection, and the absence of convalescent specimen impeded the ability to diagnose patients based on serology, particularly with respect to bacterial zoonoses. Even genuine detection of bacterial and viral DNA in normally sterile sites cannot be used to conclusively determine causality, as this has been reported among healthy individuals, and in patients even weeks after recovery from infections, challenging the interpretation of molecular assays (Williams et al., 2014;Decker et al., 1996). Blood culture was not available and this further limited the aetiological investigation. The current study did not recruit a concomitant control group, which precludes robust attribution of causality in the organisms that were detected, particularly in NP swabs. In a paediatric study in Asia and Africa, the inclusion of controls matched with pneumonia cases weakened the evidence of causality for almost all detected organisms (O'Brien et al., 2019), and only RSV, hMPV, influenza virus A and B, parainfluenza virus type 1, and B. pertussis were considered pathogenic, which is consistent with other LMIC-based studies (Bénet et al., 2017;Pretorius et al., 2016). These organisms, however, are sometimes present in healthy individuals, with a prevalence of influenza virus among healthy children between 0-6%, RSV at 0-9% and hMPV between 0-7% (Bosch et al., 2013). On the other hand, the current study did not regard other organisms detected in NP swabs such as rhinovirus, parainfluenza virus or S. pneumoniae as pathogenic, because these are commonly detected among healthy individuals (O'Brien et al., 2019;Jansen et al., 2011;Birger et al., 2018). All these limitations in reference diagnostic tests might explain the average performance of CRP in the current analysis.

Conclusion
This study presented the key organisms detected among febrile children and adults attending primary health care in Southeast Asia. The performance of CRP in distinguishing between bacterial and viral organisms was limited, although the findings suggest that CRP-guided treatment would increase the appropriate use of antibiotics with respect to aetiology. This is supported by the overall reduction in prescribing compared with current practice demonstrated in the original trial. The findings also support conclusions from previous studies that even in the presence of bacterial organisms, very few ambulatory patients are likely to benefit from the extensive and poorly targeted antibiotic prescribing practices that currently prevail in most Southeast Asian primary care settings.

Declaration of interest
The funders had no role in study design, data collection, data interpretation or writing the manuscript. The corresponding author had full access to all the data and took the final decision to submit for publication.