Article

Derivation of a clinical prediction rule to predict hospitalization for influenza in EDs

a b s t r a c t

Background: Early, rapid, and accurate identification of those patients who have severe influenza is important for emergency physicians. Influenza viral load, which has been proposed as a predictor of severe influenza, could be useful in facilitating decision making of resource use. We aimed to derive a clinical prediction rule to indicate probability for inpatient hospitalization for patients with influenza, which includes influenza viral load in addition to other clinical information commonly collected in the emergency department (ED).

Methods: We conducted a 3-year prospective cohort study (2007-2009) of patients with probable Influenza infection as suspected by the emergency physician from 3 study sites. Eligible patients were those with excess nasopharyngeal aspirate samples. Influenza viral load was measured using reverse transcription polymerase chain reaction and electrospray ionization mass spectrometry. Clinical information including demographics, underlying illness, vaccination history, hospitalization, and results from clinical laboratory were abstracted from electronic patient records and questionnaires. The prediction rule for hospitalization was derived by the recursive partitioning algorithm (decision tree-type approach) and evaluated by internal 10-fold cross- validation for performance characteristics.

Results: Of 424 ED patients with nasopharyngeal aspirates, 146 infected with influenza were enrolled (median age, 10 years [interquartile range, 4-26]; race, 55% African American; median inpatient length of stay, 3 days [interquartile range, 1-4]; high viral load group [defined as N 2.5 million genome copies/mL], 34%). Predictors for hospitalization included underlying illness, age, influenza viral load level, and vaccination history (c statistics, 0.84; sensitivity, 83%; specificity, 76%).

Conclusions: The clinical prediction rule incorporating influenza viral load into the clinical information was indicative of hospitalization and merits further evaluation for determination of ED resource use for patients with influenza.

(C) 2013

Introduction

Approximately 20% of children and 5% of adults are infected with influenza worldwide each year [1]. Influenza infection is generally self-limited but can cause severe morbidity and mortality in patients with certain risk factors, for example, extreme age, pregnancy, and high-risk medical conditions [2,3]. Early, rapid, and accurate identi- fication of those patients who have severe influenza And are at risk for a worse prognosis could help in developing effective guidelines for

? This manuscript has been presented at Society of Academic Emergency Medicine annual meeting; Phoenix, AZ; June 2010.

?? The research grant was supported by Middle Atlantic RCE Program, National

Institute of Allergy and Infectious Diseases/National Institutes of Health (5U54AI057168) and Chang Gung Memorial Hospital (CMRPCMRPG2B0271).

* Corresponding author.

E-mail address: [email protected] (K-F. Chen).

appropriate disposition and accordingly could decrease the likelihood of complications.

The emergency department (ED) is one of the most frequent health care sites to which patients with influenza-like illnesses first present. Emergency departments in the United States are increasingly over- burdened and may be poorly prepared to deal with surges in patient volumes associated with disease outbreaks [4]. The ability to quickly determine which patient with influenza/influenza-like illness should be further investigated is a complex task requiring expert clinical judgment and integration of multiple sources of data for appropriate decision making. Influenza viral load, which has been proposed to be a relatively accurate predictor for Symptom severity, inpatient length of stay, and the clinical course of infection [5,6], could serve as an ancillary biomarker to aid decision making regarding resource use in EDs.

Our goal was to derive a clinical prediction rule for inpatient hospitalization among ED patients presenting with laboratory-

0735-6757/$ – see front matter (C) 2013 http://dx.doi.org/10.1016/j.ajem.2012.10.010

confirmed influenza in this prospective ED-based observational study. Potential predictor variables included demographics, underlying illnesses, clinical findings, and influenza viral load obtained by a novel reverse transcription polymerase chain reaction and electrospray ionization mass spectrometry assay.

Materials and methods

Study design, setting, and selection of participants

Patients who visited EDs in 3 study sites of John Hopkins medical institute (site A: an Inner city Tertiary teaching hospital with 1085 beds and approximately 90 000 annual ED visits, site B: a suburban tertiary hospital with 346 beds and approximately 50 000 annual ED visits, and site C: a community hospital with 227 beds and approximately 70 000 annual ED visits) who received nasopharyngeal aspirate testing for suspected influenza ordered at the discretion of the clinicians were considered eligible during periods of December 2007 to August 2009 (sites A and C) and January to May 2009 (site B). All eligible patients were consecutively approached by dedicated study coordinators, who reviewed nasopharyngeal aspirate test order lists from the EDs daily and sought informed consent (either in person for those who were still in the hospital or by telephone if valid telephone numbers available for those who were already discharged). At least 3 separate attempts were made in different periods of different days to contact potentially eligible patients within 4 weeks after their ED visits. After written or verbal informed consent, patients who provided permission to review their medical records for microbiologic and clinical data and use of the excess nasopharyngeal aspirate were enrolled. Patients with nasopharyngeal aspirate of insufficient volume for the influenza quantification assay or those with negative influenza test results were excluded. Patients who received Antiviral medications before their ED visits were excluded, as well. The study was approved by the institutional review board at John Hopkins University.

Methods of measurement

Sample preparation

Samples were maintained in the clinical virology laboratory, frozen at -80?C, after standard virologic procedures were completed as part of their standard laboratory protocol. Samples were then processed for total nucleic acid extraction using the Thermo King- Fisher robot (Waltham, MA) according to an Ambion MagMAX viral extraction protocol (ABI, Foster City, CA). All samples were processed by a dedicated investigator (K.-F.C.) who was masked to the patient identification and results of clinical virology at the time of processing samples.

Reverse transcription PCR

One-step RT-PCR was performed in a 50-uL reaction mix consisting of 4 U of AmpliTaq Gold (Applied Biosystems, Foster City, CA); 20 mM Tris, pH 8.3; 75 mM KCl; 1.5 mM MgCl2; 0.4 Mbetaine; 800 mM mix of dATP, dGTP, dCTP, and dTTP (Bioline USA Inc, Randolph, MA); 10 mM dithiothreitol; 100 ng sonicated polyA DNA (Sigma Corp, St Louis, MO); 40 ng random hexamers (Invitrogen Corp, Carlsbad, CA); 1.2 U Superasin (Ambion Corp, Austin, TX); 400 ng T4 gene 32 protein (Roche Diagnostics Corp, Indianapolis, IN); 2 U Superscript III (Invitrogen Corp.); 20 mM sorbitol (Sigma Corp); and 250 nM of each primer. The following RT-PCR cycling conditions were used: 60?C for 5 minutes, 4?C for 10 minutes, 55?C for 45 minutes, and 95?C for 10 minutes, followed by 8 cycles of 95?C for 30 seconds, 48?C for 30 seconds, and 72?C for 30 seconds, with the 48?C annealing temperature increasing 0.9?C each cycle. The PCR was then continued for 37 additional cycles of 95?C for 15 seconds, 56?C for 20 seconds,

and 72?C for 20 seconds. The RT-PCR cycle ended with a final extension of 2 minutes at 72?C, followed by a 4?C hold.

Respiratory virus surveillance panel

The assay was performed using the Ibis T5000 Respiratory Virus Surveillance II kit (Ibis Biosciences, Inc, Carlsbad, CA), designed to detect viruses from seven groups: “conventional viruses” (respiratory syncytial virus, influenza A and B, parainfluenza types 1-4, adenovir- idae) and viruses not conventionally identified (Coronaviridae, human bocavirus, and human metapneumovirus).

Mass spectrometry for base composition analysis and quantification Reverse transcriptase PCR products were analyzed using the Ibis T5000 universal biosensor platform (Ibis Biosciences, Inc), which performs automated post-PCR desalting, electrospray ionization mass spectrometry signal acquisition, spectral analysis, and data reporting with the ideal 6-hour turnaround time, as described previously [7-10], and has the capability to approximate the quantity of pathogen within a set range. Briefly, steps were as follows: each PCR reaction was desalted and purified using a weak anion exchange protocol, as described elsewhere [11]. Accurate mass (61 ppm) and high-resolution (M/dM 100 000 full-width half-height maximum) mass spectra were acquired for each sample using high-throughput electrospray ionization mass spectrometry protocols described previously [12]. For each sample, approximately 1.5 uL of analyte solution was consumed during the 74-second spectral acquisition. Raw mass spectra were postcalibrated with an internal mass standard and deconvolved to monoisotopic molecular masses. Unambiguous base compositions were derived from the exact mass measurements of the complementary single-stranded oligonucleotides. Quantitative results were obtained by comparing the peak heights with an internal PCR calibration standard present in every PCR well at 100 molecules. Influenza viral load were measured as genome copies per milliliter and categorized into tertiles (high, medium, and low) in the study owing to no consistent threshold available at this point. Health care providers (HCPs) were masked of the

viral load information during the care of patients.

Data collection

Clinical information was obtained by medical record review to obtain patient demographic (ie, age, sex, race, and study site), medical history (eg, any underlying illness such as diabetes, chronic lung disease defined as asthma or chronic obstructive pulmonary disease, and immunocompromised status defined as currently under chemotherapy or corticosteroids), symptoms and signs (eg, days of illness, fever, cough, and shortness of breath as recorded in medical records), diagnostic tests performed in the ED and their findings (ie, laboratory such as white blood count and differential counts, hemoglobin, vital signs such as body temperature at triage, and abnormal chest radiographs defined as infection related findings, eg, infiltration), ED triage level, and clinical outcomes (ie, hospitalization, admission to intensive care unit, and length of stay in hospital in days if admitted), as well as interview in person or by telephone to obtain influenza vaccination history (seasonal influenza vaccination received before the season) and sick contact information. Clinical laboratory results and vital signs were categorized into higher, lower, or within the reference range, when appropriate; determined by different age criteria according to the clinical pathology in Johns Hopkins University; and coded as not tested or recorded if tests were not ordered or recorded by HCPs.

Statistical methods

Recursive partitioning algorithm (rpart packages of the R statis- tical analysis program) was applied to obtain the prediction rule for hospitalization. Recursive partitioning algorithm used a 2-stage procedure to derive binary trees. The classification tree was built in

the first stage by finding the best variables to separate data to maximize purity by using Gini index in each node recursively until no improvement can be made. The optimal cut points of the continuous variables were found automatically by the algorithm to build the binary tree. Surrogate variables were found by the algorithm for variables with missing data to “best utilize” the predictors without discarding any patients with missing data [13]. The candidate classification trees were evaluated by the minimal relative error rule by cost-complexity parameter using internal 10-fold cross-validation. Briefly, the data were divided into 10 equal parts randomly, and then the classification tree was derived with 9 parts (the learning data set) and tested with 1 part (the validation data set). The 10-fold cross- validation was repeated 10 times, and the results were combined to develop the relative error for the tree. Another 5-fold cross-validation was performed to acquire the average cross-validated performance characteristics (ie, accuracy, which was defined as proportion of true results [both true positives and true negatives]; sensitivity; specific- ity; and c statistics) of the derived and pruned classification tree. Similarly, the data were divided into 5 equal parts randomly, and the predicted admissions by classification tree were compared with the observed level in one part sequentially. The average performance characteristics were then obtained from the 5 parts.

Binary or categorical variables were tested by ?2 or Fisher exact

test, when appropriate, and continuous variables were tested by Wilcoxon rank sum test or Kruskal-Wallis equality-of-populations rank test. A subgroup analysis was performed for patients who visited the different study sites because the practice of emergency physicians could be different. To assess the consistency of the results in different study sites, the association between hospitalization and different study sites was tested, and the clinical prediction rule was evaluated for different study sites. Association between influenza viral load and hospitalization for different underlying illness status was evaluated. Analyses were performed using R free software (version 2.11; R Foundation for Statistical Computing, Vienna, Austria) and Stata software (version 11.1; Stata Corp, College Station, TX). A 2-tailed P value less than .05 was considered significant.

Results

Overall recruitment

During the 21 months of recruitment period, 1188 patients were eligible and 424 (36%) consented in this study (Fig. 1). Most of the patients who were not recruited were those who could not be contacted because of wrong telephone number or failure to contact (n = 725; 95%). Among the 160 patients with influenza viral infection confirmed by the clinical virology laboratory, 14 who had nasopha- ryngeal aspirate testing during hospitalization were excluded. Of the remaining 146 patients, 100 (69%) were younger than 18 years, 79

(55%) were African Americans, 79 (55%) were male, 117 (81%) were recruited in the study site A, and most patients had at least 1 underlying illness (69% [n = 99]; Table 1).

Clinical findings and outcomes

Approximately most patients presented to the ED with less than or equal to 3 days of illness (63%; n = 92), and the most commonly reported symptoms were fever and cough (35% and 31%, respective- ly). Upon presentation to the ED, most patients were categorized into lower Acuity levels of triage, that is, in level 4 or 5 (59%). More than half of the patients had tachycardia (78%) or tachypnea (60%). More than one-third of the patients (n = 56; 38%) were admitted to the hospital, with median inpatient length of stay of 3 days. Among one- third of patients who had chest radiographs performed (including chest plain film and computed tomography), 41% had positive pulmonary infection related findings, for example, pulmonary

Fig. 1. Recruitment diagram. During the 21 months of recruitment period, 424 of 1188 patients consented in this study. After further excluding 264 non-influenza-infected patients and 14 patients with NPA collected not at EDs, 146 patients were included in the final analysis. Influenza positive: positive by clinical virology laboratory. NPA, nasopharyngeal aspirate.

infiltration. Among the 89 (60%) patients who received white blood cell count tests, nearly half (57%) had findings within the reference range (Table 1). Patients who were vaccinated for seasonal influenza tended to have influenza infection history in the past 2 years, history of asthma, or have immunocompromised status (P b .01, .07, and .03, respectively). Patients with any underlying illness (69%; n = 99) were more likely to be vaccinated against seasonal influenza (43%) than healthy patients (24%; ?2 test, P = .03).

Reverse transcriptase PCR and electrospray ionization mass spectrometry virology testing

The average (SD) storage time of nasopharyngeal aspirates before processed in RT-PCR and electrospray ionization mass spectrometry was 7 (1.4) months. Influenza virus was not detected by RT-PCR and electrospray ionization mass spectrometry platform in 4 of the clinical virology laboratory influenza-positive samples, and “zero” viral load was assigned to them. The median viral load was approximately 4.5 x 105 copies/mL, with interquartile range of 9.8 x 104 to 1.7 x 106 copies/mL. Type A influenza viruses were found in 73 (62%) specimens, among which 53 were seasonal H1N1, 11 were H3N2, 7 were found to be pandemic influenza A (H1N1) 2009, and 2 were H2N2. Viral load from samples that were positive for pandemic H1N1 group were not different from the seasonal H1N1 group (Wilcoxon rank sum test, P = .93); therefore, no further stratification analysis by H1N1subgroup was performed to derive the clinical decision rule. Forty patients (34%) were categorized into high viral load group (>=2.5 million genome copies/mL). However, influenza viral loads among those hospitalized patients were not statistically significant higher than those patients discharged from the EDs (data not presented).

Characteristics of hospitalized patients

Patients who were hospitalized tended to have at least an underlying illness (89% vs 56% [inpatient vs outpatient for successive respective values]; Table 1), chest radiograph testing (54% vs 27%), white blood cell count (91% vs 61%), absolute neutrophil cell count (75% vs 23%), absolute lymphocyte cell count (75% vs 30%), hemoglobin testing (61% vs 28%), a prescription of an antiviral agent in ED or hospital (20% vs 0%), and positive chest radiographic finding (53% vs 25%) (all P b .05). However, patients who were hospitalized had similar age, sex and race distribution, days of illness

Table 1

Patient characteristics

Characteristics Whole population (N = 146)

Inpatients (n = 56)

internal cross-validation, the tree was trimmed to have 5 nodes. Four predictors chosen by the rpart algorithm were any underlying illness, age, influenza viral load level, and vaccination history; 2 different cutoff points of age were obtained in the final decision classification

Demographics Median age (y) 10 (4-26) 11.5 (2.8-27.5)

Male 79 (55) 28 (50)

African American 79 (55) 33 (59)

Study site A 117 (81) 52 (93)

Study site B 5 (3) 2 (3)

Study site C 22 (15) 3 (5)

tree (Fig. 2). For patients with any underlying illness, being younger than 7 months (100% hospitalized) or older than 14 years (64%), or being vaccinated against seasonal influenza in the past 2 years (67%) predicted admission to the hospital. For patients who did not have any underlying illness, having high influenza viral load was predictive of

Medical history

Any underlying illness 99 (69) 50 (89)?

Chronic lung disease 22 (15) 8 (14)

Immunocompromised 18 (13) 9 (16)

Influenza vaccination 54 (38) 23 (41)

Sick contact 46 (32) 17 (30)

hospitalization (100% hospitalized). The classification tree derivation was cross-validated to have performance of a sensitivity of 83% (73%- 90%), a specificity of 76% (95% confidence interval [CI], 63%-86%), and

a c statistic of 0.84 (0.77-0.90).

Symptoms Median days of illness 3 (2-4)a 3.5 (2-5)a Fever 50 (35) 15 (27)

Cough 45 (31) 15 (27)

Sputum 6 (4) 4 (7)

Rhinorrhea 18 (13) 8 (14)

Sore throat 5 (3) 2 (4)

Shortness of breath 9 (6) 7 (13)?

GI symptomsb 22 (15) 7 (13)

Headache 7 (5) 0 (0)?

General achiness 11 (8) 1 (2)?

General malaise 5 (3) 4 (7)

Chest discomfortc 6 (4) 3 (5) Vital signs Hyperthermia 35 (34)a 9 (24)a

Tachycardia 80 (78)a 31 (84)a

Tachypnea 61 (60)a 24 (75)a

Hypotension 4 (4)a 2 (6)a

3.6. Subgroup analysis

Patients who visited the study site A were more likely to be hospitalized (44% vs 20% and 14% compared with sites B and C, same expression as below) and have any underlying illness (74% vs 60% and 45%) but less likely to receive absolute lymphocyte and neutrophil Cell differential count testing (45% vs 80% and 86%) and chest radiographs (58% vs 20% and 95%) (All P b .05). Nevertheless, patients at different study sites had similar distribution of age, sex and race, days of illness before ED visits, influenza vaccination history, and influenza viral load levels. In light of the different characteristics of patients visiting study site A, the sensitivity analysis was performed by evaluating the

classification tree among site A patients. The classification tree has

Clinical

finding

Clinical

Triage level b 3 24 (59)a 6 (67)a Chest radiograph result

Positive 22 (15) 16 (29)?

Negative 32 (22) 14 (25)?

Not tested 90 (63) 26 (46)? White blood cell count

High 8 (6) 5 (9)?

Normal 51 (35) 32 (57)?

Low 28 (19) 14 (25)?

Not tested 57 (40) 5 (9)? Absolute Neutrophil count

High 17 (12) 14 (25)?

Normal 36 (25) 22 (39)?

Low 13 (9) 6 (11)?

Not tested 78 (54) 14 (25)? Absolute lymphocyte count

High 1 (1) 0 (0)?

Normal 27 (19) 18 (32)?

Low 40 (28) 24 (43)?

Not tested 76 (53) 14 (25)? Hemoglobin level

High 3 (2) 3 (4)?

Normal 27 (19) 16 (29)?

Low 29 (20) 16 (29)?

Not tested 85 (60) 22 (39)?

Hospitalization 56 (39) –

performance of a sensitivity of 79% (95% CI, 72%-93%) and specificity of 71% (83%-97%) among site A patients, which are not significantly different from the overall performance (both P N .05). Higher influenza viral load (>=2.5 million genome copies/mL) was found to be associated with hospitalization for patients without any underlying illness (Fig. 2; odds ratio, 14.5; 95% CI, 1.5-139.5; P = .012). Furthermore, this association was found to be stronger for patients recruited at study site A (odds ratio, 20; 95% CI, 1.9-211.8; P = .013).

Discussion

This prospective ED-based study that included mostly pediatric patients with underlying illness demonstrated that an accurate classification prediction tree could be derived to predict hospitaliza- tion, which offered a potential adjunctive guide for primary HCPs for decision making regarding the need for further investigation for influenza-infected patients. Need for hospitalization, as a surrogate of severe influenza viral infection, could be predicted by underlying illness, age, influenza viral load levels, and vaccination history with good performance characteristics.

Influenza viral load has been studied in several different popula-

outcome

Median inpatient length of stay (d)

3 (1-4) –

tions and has been shown to be controversially associated with symptom severity, disease severity, patient comorbidity, disease

ICU admission 7 (5) 7 (13)

Received antivirals later 11 (8) 11 (20)?

Values are presented as IQR or n (%). Chronic lung disease: asthma or chronic obstructive pulmonary disease. Percentages may not add to 100% due to rounding.

ICU, intensive care unit; GI, gastrointestinal.

* P b .05.

a Numbers of missing values found.

b Gastrointestinal symptom included abdominal pain, nausea, and vomiting.

c chest discomfort included chest pain, achiness or any chest complaint.

before ED visits, symptoms and vital signs profiles, and influenza viral load levels compared with patients discharged from the EDs (Table 1).

3.5. Classification tree derivation

Using the rpart algorithm, we developed a full-grown tree with 13 nodes (not shown). According to the rule of minimal relative error by

course, and patient age [5,14]. Although Lee et al [5] found higher viral load among inpatients population, viral load was found not predictive of hospitalization in a study population with milder disease in another study [15]. In our study, we found that higher viral load was predictive of hospitalization for patients without any underlying illness. We speculate, therefore, that initial influenza viral load could have much more predictive value for the need of hospitalization in certain patient populations (eg, milder influenza infection or healthier population) and that clinicians could utilize this biomarker to allocate limited resource in the ED.

Not surprisingly, infected patients with an underlying illness were likely to be admitted to hospitals more frequently than those without any underlying illness. As found in our classification trees, underlying illness was always the most influential predictors of hospitalization, which was treated as the first and most important prediction node by recursive partitioning algorithm. Most of the studies and guidelines

Fig. 2. Classification tree to predict hospitalization. This tree-type clinical prediction rule with any underlying illness, age, vaccination history, and influenza viral load was derived by the recursive partitioning algorithm. The most important indicator for hospitalization was underlying illness, as indicated as the first node in this tree. For those patients without any underlying illness, high viral load would predict hospitalization (N 2.5 million copies/mL). Extreme of age, that is, younger than 7 months or older than 14 years, was predicted of hospitalization for those patients. Vaccination history for those patients with underlying illness and aged between 7 months and 14 years was predicted of hospitalization. Percentages of admitted patients were presented by each decision box. VL, viral load.

recommended admitting patients with “high-risk medical condi- tions,” and our prediction rule also demonstrated that current practice was consistent with the guideline, in both tertiary teaching hospitals and the community hospital. Whether or not admitting all infected patients with any underlying illness is cost-effective still remains unsolved and further study is merited.

Age was found to be predictive of admission for patients with influenza in our study, which is consistent with multiple prior studies reporting that the very young patients (ie, b 6 months) were most likely to be admitted [16-19]. Accordingly, age could be an important predictor of severe influenza, indicating that HCP should be aware and make disposition for influenza-infected patients accordingly.

Counterintuitively, patients with influenza vaccination were found to have higher probability of being admitted in our study. We also found that those patients were more likely to have had a history of influenza in the past 2 years, a history of asthma, or an immunocom- promised status. We speculated that vaccination history may be an indicator of a group of risk factors, which were associated with increased likelihood of being vaccinated and were accordingly treated as high risk and more likely to be admitted.

Classification or regression trees obtained by recursive partitioning algorithm presented as the hierarchical trees, as shown in our study, are believed to be close to clinical thinking process, which could provide useful tools to aid clinical daily practices. However, interpretation and explanation of the trees can be difficult because each predictor, except the first node, is interacting with previous higher hierarchical predictors. Because some researchers recommend using tree-based approaches as the tool for data mining [20], we caution readers to keep in mind this complicated interacting relationship. The optimal use of the tree could be in combination with bivariate data analysis, or via incorporating previous evidence.

Unfortunately, this approach could not be taken here mainly because of missing values.

Limitations

Our study has several important limitations. First, many missing values were found in our final data set. We speculate that these were not missed at random, especially for those missed laboratory tests. However, the recursive partitioning algorithm has the advantage to manage observations with missing values by using surrogate splits, which might mitigate this limitation [21]. Second, the eligibility of our study population was at the discretion of the ED physicians, that is, testing of nasopharyngeal aspirate, which might have led to selection bias in the study population. However, we aimed to apply this clinical prediction rule to patients whom clinicians thought nasopharyngeal aspirate testing was warranted; therefore, selection bias might not be an issue. Third, this prediction rule was derived from laboratory- confirmed influenza virus-infected patients; therefore, we caution readers about generalizing this prediction rule among those patients confirmed by less accurate rapid influenza antigen tests. Fourth, information obtained through telephone interview (eg, vaccination history and sick contact) might have occurred weeks after the patients visited the ED, thus resulting in some potential information bias. However, we did review the medical record thoroughly to minimize this bias. Fifth, because some have started to use quantification of cellular gene to better measure the quality of specimens and to correct influenza viral load [16], our study, which was designed to use broad- based RT-PCR and electrospray ionization mass spectrometry plat- form, has yet to build in this kind of quality assurance mechanism and could not fulfill this kind of viral load correction. Sixth, the RT-PCR and electrospray ionization mass spectrometry platform we used to obtain

viral loads in this study may not be readily available in all institutions. Although this platform had been validated for semiquantification [8,22], further studies are merited to verify the association between viral load and severity of influenza. Seventh, because of the difficulty to enroll patients during their short ED stays, we cannot access patients who had insufficient volume of specimen. However, the number of those missed patients visits was not large (b 10%, data not presented). Last is the generalizability issue; because the primary study site is a Tertiary medical center that receive many referred cases, further studies may be merited to validate this prediction rule in different clinical settings. We also caution the readers that the most of the subjects enrolled in this study had underlying illness, which might represent sicker population. Further study should be conducted to validate our findings.

Conclusions

In conclusion, in this prospective cohort study, we demonstrated that underlying illness, age, vaccination history, and influenza viral load level could be used to predict the need for further investigation for possible hospitalization of patients with confirmed influenza in the ED. Further prospective study is merited to externally validate this prediction rule.

References

  1. Nicholson KG, Wood JM, Zambon M. Influenza. Lancet 2003;362:1733-45.
  2. Thompson WW, Shay DK, Weintraub E, et al. Mortality associated with influenza and respiratory syncytial virus in the United States. JAMA 2003;289: 179-86.
  3. Irving WL, James DK, Stephenson T, et al. Influenza virus infection in the second and Third trimesters of pregnancy: a clinical and seroEpidemiological study. BJOG 2000;107:1282-9.
  4. Rothman RE, Irvin CB, Moran GJ, et al. Respiratory hygiene in the emergency department. J Emerg Nurs 2007;33:119-34.
  5. Lee N, Chan PK, Hui DS, et al. Viral loads and duration of viral shedding in adult patients hospitalized with influenza. J Infect Dis 2009;200:492-500.
  6. Leekha S, Zitterkopf NL, Espy MJ, et al. Duration of influenza A virus shedding in hospitalized patients and implications for Infection control. Infect Control Hosp Epidemiol 2007;28:1071-6.
  7. Sampath R, Hofstadler SA, Blyn LB, et al. rapid identification of emerging pathogens: coronavirus. Emerg Infect Dis 2005;11:373-9.
  8. Grant RJ, Baldwin CD, Nalca A, et al. Application of the Ibis-T5000 pan- Orthopoxvirus assay to quantitatively detect monkeypox viral loads in clinical specimens from macaques experimentally infected with aerosolized monkeypox virus. Am J Trop Med Hyg 2010;82:318-23.
  9. Chen KF, Rothman RE, Ramachandran P, et al. Rapid identification viruses from nasal pharyngeal aspirates in acute viral Respiratory infections by RT-PCR and electrospray ionization mass spectrometry. J Virol Methods 2011;173:60-6.
  10. Chen KF, Blyn L, Rothman RE, et al. Reverse transcription polymerase chain reaction and electrospray ionization mass spectrometry for identifying acute viral upper respiratory tract infections. Diagn Microbiol Infect Dis 2011;69:179-86.
  11. Jiang Y, Hofstadler SA. A highly efficient and automated method of purifying and desalting PCR products for analysis by electrospray ionization mass spectrometry. Anal Biochem 2003;316:50-7.
  12. Sampath R, Russell KL, Massire C, et al. Global surveillance of emerging Influenza virus genotypes by mass spectrometry. PLoS One 2007;2:e489.
  13. Zhang H, Singer B. Recursive partitioning in the health sciences. New York: Springer-Verlag; 1999.
  14. Sato H, Carvalho G, Sato T, et al. The association of preoperative Glycemic control, intraoperative insulin sensitivity, and outcomes after cardiac surgery. J Clin Endocrinol Metab 2010;95:4338-44.
  15. Nilsson AC, Brytting M, Serifler F, et al. Longitudinal clearance of seasonal influenza A viral RNA measured by real-time polymerase chain reaction in patients identified at a hospital emergency department. Scand J Infect Dis 2010;42:679-86.
  16. Quach C, Piche-Walker L, Platt R, et al. Risk factors associated with severe influenza infections in childhood: implication for vaccine strategy. Pediatrics 2003;112:e197-201.
  17. Izurieta HS, Thompson WW, Kramarz P, et al. Influenza and the rates of hospitalization for respiratory disease among infants and young children. N Engl J Med 2000;342:232-9.
  18. Thompson WW, Shay DK, Weintraub E, et al. Influenza-associated hospitalizations in the United States. JAMA 2004;292:1333-40.
  19. Li C-CWL, Eng H-W, You H-W, Chang L-S, Tang K-S, et al. Correlation of pandemic (H1N1) 2009 viral load with disease severity and prolonged viral shedding in children. Emerg Infect Dis 2010;16:1265-72.
  20. Hastie TT R, Friedman J. The elements of statistical learning, chapter 10. 2nd ed. New York: Springer; 2009.
  21. Hapfelmeier A, Hothorn T, Ulm K. Recursive partitioning on incomplete data using surrogate decisions and multiple imputation. Computat Stat Data Anal 2012;56: 1552-5.
  22. Hofstadler SA, S.R., Blyn LB, Eshoo MW, Hall TA. TIGER: the universal biosensor. Int J Mass Spectrom 2005;242:23-41.

Leave a Reply

Your email address will not be published. Required fields are marked *