Accuracy of individual signs and symptoms and case definitions for the diagnosis of influenza in different age groups: a systematic review with meta-analysis

Abstract

Objective

The objective of this study is to perform a comprehensive systematic review and meta-analysis of the accuracy of signs, symptoms and case definitions for the diagnosis of influenza.

Design

Systematic review and meta-analysis of diagnostic accuracy.

Setting

Inpatient or outpatient setting.

Participants

Three databases (PubMed, CINAHL and EMBASE) were searched through February 2024 for studies of clinical diagnosis of influenza using prospective data collection and a high-quality reference standard. Data were abstracted by researchers working in parallel and resolving discrepancies by discussion.

Primary and secondary outcome measures

Quality was assessed using QUADAS-2. Summary estimates (or ranges) of sensitivity and specificity, likelihood ratio (LR), the Youden Index and the area under the receiver operating characteristic curve were calculated.

Results

The final meta-analysis included 67 studies, each with between 119 and 155 866 participants. Most were judged to be low risk of bias. The signs and symptoms with the highest overall accuracy for all studies based on the Youden Index were any fever (0.32), overall clinical impression (0.28), coryza (0.25), cough and fever (0.25), and measured fever (0.25). Accuracy varied widely by age group. Only the overall clinical impression had a positive LR greater than 2.0. Cough was the most sensitive finding (0.92) with a negative LR of 0.28 in adults. The absence of any fever also had a low negative LR (0.30). The Centers for Disease Control and Prevention (CDC) definition of influenza-like illness (ILI) had good specificity but poor sensitivity in adults, while in infants, it had good sensitivity but widely varying specificity. The European CDC and WHO case definitions for ILI had modest sensitivity and specificity.

Conclusions

Individual signs and symptoms, their combinations, and ILI case definitions have very limited accuracy for identifying persons with influenza. More accurate surveillance and diagnosis will require the development and validation of accurate risk scores or greater use of point-of-care testing.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This was a comprehensive search of the literature through 2024 that identified 67 studies.

• Bias was assessed using the QUADAS-2 framework; and the included studies all used prospective data collection and a highly accurate reference standard test.

• Analysis used modern methods for meta-analysis of diagnostic accuracy.

Introduction

Seasonal influenza accounts for an average of 389 000 deaths worldwide every year and has recurred after the COVID-19 pandemic as people have reduced their use of social distancing and mask-wearing.^{1 2} Neuraminidase inhibitors, such as zanamivir and oseltamivir, can reduce the duration and severity of influenza symptoms for both influenza virus type A and B but must be initiated within 36 hours of symptom onset to be effective.^{3 4} Prompt diagnosis of influenza is, therefore, needed to avoid overuse of antiviral therapy and to recommend infection control measures. While rapid antigen and molecular tests for influenza are widely used in some countries such as the USA,⁵ in most countries and especially in low-resource settings, diagnosis relies on interpretation of clinical signs and symptoms.⁶

The World Health Organization (WHO), the US Centers for Disease Control and Prevention (CDC) and others have developed clinical case definitions for influenza-like illness (ILI) that are used both for disease surveillance and for clinical decision-making. These same organisations have also created definitions for acute respiratory infection (ARI) and severe ARI (SARI) that are also used clinically and for disease surveillance (table 1). To date, though, the diagnostic accuracy of these definitions has not been assessed by systematic reviews and meta-analyses.

Table 1. Case definitions for influenza-like illness and acute respiratory illness.

Source	Definition
Influenza-Like Illness (ILI)
Centers for Disease Control and Prevention Influenza Like Illness (https://www.cdc.gov/flu/weekly/overview.htm)	CDC_ILI: Measured fever AND (cough OR sore throat) and no known cause other than influenza
WHO 2018Influenza Like Illness11	WHO_ILI: Cough AND measured fever AND onset in last 10 days
European Centers for Disease Control Influenza-Like Illness20	ECDC_ILI: At least one of: (measured fever, feverishness, headache, malaise or myalgia) AND at least one of (cough, sore throat or shortness of breath)
Acute Respiratory Illness (ARI)
European Centers for Disease Control Acute Respiratory Illness20	ECDC_ARI: Acute onset and at least one of the following (cough, sore throat, shortness of breath or nasal discharge) and thought to be caused by infection
European Centers for Disease Control Febrile Acute Respiratory Illness20	ECDC_FARI: ECDC ARI with self-reported fever
WHO 2018Severe Acute Respiratory Illness11	WHO_SARI_2018: Acute respiratory infection with cough AND measured or subjective fever AND onset <10 days AND requires hospitalisation
WHO 2011Severe Acute Respiratory Illness (<5 years)11	WHO_SARI_2011_LT5: (Cough or difficulty breathing) AND (unable to feed or vomits or convulsions or lethargic/unconscious or chest indrawing or stridor) AND requires hospitalisation
WHO 2011Severe Acute Respiratory Illness (≥5 years)11	WHO_SARI_2011_GT5: Fever (measured) AND (cough or sore throat) AND (shortness of breath OR difficulty breathing) AND requiring hospital admission
Ontario Ministry of Health Febrile Respiratory Index (www.health.gov.on.ca/english/providers/program/emu/health_notices/ihn_gd_ed_042909.pdf)	Ontario_FRI: (Shortness of breath OR cough) AND (subjective fever OR measured fever OR chills)
Netherlands Medical Assistance for Accidents and Disasters (https://www.ghor.nl)	Dutch_GHOR: Measured fever AND 2 or more of cough, rhinorrhoea, sore throat, headache, myalgia, malaise, chills
Thai Bureau of Epidemiology21	TBE: Fever AND myalgia AND one of the following: headache, cough, sore throat, prostration, nasal congestion and conjunctivitis

Systematic reviews have been performed of individual signs and symptoms for influenza but were limited by the search strategy, by not using modern bivariate analytic methods for diagnostic meta-analysis, and because they were published 10 or more years ago,6,8 In addition, previous reviews did not stratify assessment of accuracy by patient age. Our goal is, therefore, to perform a comprehensive systematic review and meta-analysis of signs, symptoms and definitions of ILI, ARI and SARI for the diagnosis of influenza stratified by age group.

Methods

This was a systematic review and meta-analysis of published studies of the accuracy of signs, symptoms and ILI definitions for the diagnosis of influenza. The study was registered with the PROSPERO database (#CRD42020161801) and followed PRISMA guidance regarding conduct and reporting of a diagnostic meta-analysis.⁹

Patient and public involvement

Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research.

Inclusion criteria

Inclusion criteria were selected to avoid spectrum bias, verification bias and information bias. Studies were included if they recruited a prospective cohort of infants, children, adolescents or adults presenting with symptoms of respiratory tract infection (RTI) or clinically suspected influenza in the inpatient or outpatient setting. Studies had to report sufficient information to calculate sensitivity and specificity for the diagnosis of influenza for at least one sign, symptom, combination of symptoms or case definition such as ILI or SARI. No limits were set for country or year; language was limited to those spoken by at least one of the authors (English or German). The reference standard had to be PCR, culture or paired serology and had to have been performed in all participants to avoid verification bias.

Studies were excluded if they did not enrol patients with acute RTI or suspected influenza, or if they were in a specialised population such as only patients in skilled nursing facilities, immunosuppressed patients, or patients with chronic lung disease. Studies were also excluded if they gathered data retrospectively using chart or medical record review, or if they used a diagnostic case–control design (ie, recruited patients with known influenza and healthy controls or used banked samples).

Search strategy

The search was built around the concepts for “signs, symptoms, clinical prediction rules”, “influenza” and “diagnosis” linked by Boolean AND joins and used three databases (ie, PubMed, Embase and CINAHL) and is shown in online supplemental appendix A. The limits ‘has abstract’ and ‘English’ were applied to the search using the PubMed database only; there was no limitation by age of study. In addition, the reference lists of included studies were reviewed for additional articles, as were three older systematic reviews identified by our search. A separate search for studies of the accuracy of case definitions was also performed, and the results merged (online supplemental appendix B). Finally, a bridge search was performed in February 2024.

Data abstraction

All abstracts were reviewed for inclusion in parallel by two authors, one of whom was always a physician. For any abstract deemed potentially of interest, the full article was obtained and reviewed in parallel by two authors to determine inclusion. Studies meeting inclusion and exclusion criteria were reviewed in parallel by two authors who each abstracted study characteristics, study quality and test accuracy data. Discrepancies were resolved through consensus discussion.

Quality assessment

The QUADAS-2 framework was used for our study and definitions for low, unclear and high risk of bias prespecified for each domain. These are shown in online supplemental appendix C.¹⁰ It was assumed that ascertainment of clinical signs and symptoms was masked to the results of PCR, culture or serology performed later in a reference laboratory.

Data preparation

Similar signs and symptoms were grouped together. For example, ‘purulent sputum’, ‘sputum (purulent)’ and ‘mucopurulent sputum’ were grouped into a single variable called ‘sputum (purulent)’. For vital signs and other continuous variables, similar cut-offs to define an abnormal finding were combined where clinically reasonable, that is, temperature greater than 37.7°C and temperature greater than 38°C. For studies reporting the overall clinical impression in multiple categories of disease likelihood, categories were combined to create a dichotomous variable.

With regard to case definitions of ILI, data for the older WHO definition and for the US Department of Defense Definition was combined with that of the CDC ILI as they are essentially identical (measured fever AND [cough OR sore throat]), with the CDC adding “and no known cause other than influenza”, the WHO adding “and thought to be due to respiratory infection”, and the US Department of Defense adding “and specimen collected within 72 hours”.¹¹

We defined infants as less than 5 years, children as 5–17 years and adults as 18 years or older. Participants were classified based on the best fit, for example, a study of persons 16 and older would be classified as ‘adult’, of children 1–17 years as ‘infant and child’, and of infants 3 months to 3 years as ‘infant’.

Analytical strategy

Data were imported into R (V.4.3.2, 2023) using the R Studio framework (V.2023.12.1+402). If there were five or more studies of a sign, symptom or case definition in any age group or overall, then bivariate meta-analysis was performed and summary estimates for the sensitivity, specificity and likelihood ratios (LR) were reported, as well as the area under (AUC) the receiver operating characteristic (ROC) curve and the Youden Index. A sign, symptom or case definition was considered significantly associated with the diagnosis of influenza if the 95% CI for the LRs excluded 1.0.

The Youden Index is the sum of sensitivity and specificity minus one and is an overall measure of discrimination.¹² A value of 0.0 is equivalent to an AUC of 0.5, a value of 1.0 indicates a perfectly accurate test. Negative values suggest that a positive test (eg, presence of a sign or symptom) reduces the likelihood of influenza. The mada package (V.0.5.11) was used to calculate summary ROC curves and measures of accuracy with 95% CIs. If there were 1–4 studies of a sign or symptom in an age group or overall, then the range or point estimate of sensitivity and specificity are provided and the AUC and Youden index are not reported.

Results

Our search of the literature is summarised in figure 1. A total of 1449 unique studies were identified, and 281 were reviewed in full text. The final systematic review included 67 studies, each with between 119 and 155 866 participants (median of 838 participants). Characteristics of the 67 included studies are summarised in table 2 (a table including inclusion criteria, country and year is in online supplemental appendix table 1). Studies were set in a wide range of countries, including 27 in North America, 16 in Europe, 12 in Asia, 5 in Africa, 3 in Australia or New Zealand, 3 in the Middle East and 1 in multiple regions of the world. Where reported, the mean or median age ranged from 22 months to 76 years. The most common settings were outpatient and/or emergency department for 39 studies, followed by inpatient for 11 studies and population or school-based for 11 studies. 10 of the outpatient studies specifically described being a primary care setting (general practice, family medicine or paediatrics). With regard to the reference standard, 55 studies used PCR, 4 viral culture, 2 paired serology and the remainder a combination of these reference standards.

Figure 1. PRISMA diagram showing the results of the literature search. PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Table 2. Characteristics of included studies.

Study	# of patients	Mean age (years)	Setting	Influenza strain	Reference standard
Andrew, 202022	11 379	72 (median)	Inpatient	A+B	PCR
Aung, 202123	717	36	ED	A+B	PCR
Barbara, 201224	142	22.1	Population based	A+B	PCR
Barry, 202125	1653	12.5	Population based	A+B	PCR
Bellei, 200726	343	34 (median)	Outpatient and healthcare worker	A+B	PCR
BinSaeed, 201127	184	25	Outpatient	H1N1	PCR
BinSaeed, 201428	290	25.2	Outpatient	H1N1	PCR
Brouqui, 200929	307	NR	Outpatient	A	PCR
Bruning, 201830	353	45 (median)	Outpatient	A	PCR
Bryant, 201031	446	NR	Outpatient	A	PCR
Campe, 201632	665	NR	Outpatient primary care	A+B	PCR
Carrat, 199933	610	NR	Outpatient primary care	A+B	PCR+
Casalegno, 201734	14 994	18.5	Outpatient primary care	A+B	PCR
Conway, 201335	944	22 months (median)	Outpatient, inpatient and ED	A+B	PCR and culture
Crum-Cianflone, 200936	697	20 (median)	Outpatient	A+B	PCR
Davis, 202237	6378	NR	Inpatient	A+B	PCR
DeMarcus, 201838	5575	NR	Outpatient	A+B	PCR
Dugas, 201539	270	50	ED	A+B	PCR
Dugas, 201940	1941	48.6	ED	NR	PCR
Duque, 201041	828	43 (median)	Inpatient and outpatient	A	PCR
Farrell, 201342	217	NR	Outpatient	A	PCR
Gupta, 201243	1043	NR	Inpatient	A	PCR
Guzmán-Esquivel, 202344	6027	35.2	Outpatient	A+B	PCR
Hirve, 201245	3179	NR	Inpatient	A+B	PCR
Hombrouck, 201246	949	NR	Outpatient primary care	A+B	PCR
Hsieh, 201447	239	NR	School based	A+B	Paired serology
Hulson, 200148	432	28.9	Outpatient primary care	NR	Culture
Hung, 202349	2189	40 (median)	Inpatient and ED	A+B	PCR
Jiang, 201550	700	44	Population based	A	Paired serology
Kasper, 201051	2639	11 (median)	Outpatient	A+B	PCR
Kim, 201052	828	32 (median)	Outpatient	A	PCR
Kotnik, 202253	976	36 (median)	Population based	A+B	PCR
Lam, 201654	1318	76.4 (median)	ED	A+B	PCR
Maman, 201455	955	15.5	Outpatient	A+B	PCR
Michiels, 201156	4584	30	Population based	A+B	PCR
Monamele, 202057	11 816	NR	Outpatient	A+B	PCR
Monto, 200058	3744	34.7	Outpatient	A+B	PCR+
Moretti, 201159	513	27.9	Inpatient	A+B	PCR
Mulpuru, 201360	200	66	Inpatient	NR	PCR
Murray, 201361	4800	12	Inpatient	A+B	PCR
Neuzil, 200362	585	67.6	Population based	A+B	Culture
Ngobeni, 201963	1939	NR	Inpatient	A+B	PCR
Nitsch-Osuch, 201364	150	NR	Outpatient	A+B	PCR
Okiyama, 2022 (deriv)13	7831	33.8	Inpatient or outpatient	A+B	PCR
Okiyama, 2022 (valid)13	659	33.8	Inpatient or outpatient	A+B	PCR
Ouchi, 202265	2614	41.8	Outpatient primary care	A+B	PCR
Padin, 201466	21 570	21	Outpatient	A	PCR
Pedersen, 201967	119	7 (median)	ED	A+B	PCR
Prasarnphanich, 201021	838	NR	Outpatient	A+B	Culture
Puzelli, 200968	580	24 (median)	Outpatient primary care	A+B	PCR
Radin, 201469	1444	NR	Population based	A+B	PCR
Rao, 202270	1478	3.2	ED and outpatient	A+B	PCR
Rowlinson, 202171	6113	NR	Inpatient	A+B	PCR
Senn, 200572	201	34.3	Outpatient primary care	A+B	Culture
Shahid, 201073	315	74	Population-based	A+B	PCR
Smit, 201174	964	36	Outpatient	A	PCR
Sreng, 201014	155 866	NR	Inpatient and outpatient	A+B	PCR
Stein, 200575	258	38	ED or outpatient	A+B	PCR
van den Dool, 200876	264	60	Inpatient	NR	PCR
van Vugt, 201577	1801	48	Outpatient primary care	A+B	PCR
Vuichard-Gysin, 201978	2191	NR	Population based	A+B	PCR
Walsh, 200279	332	77	Inpatient	A	PCR+
Woolpert, 201280	789	31.2	Outpatient and ED	A+B	PCR
Yang, 201581	158	35 (median)	Outpatient	A+B	PCR and/or culture
Zambon, 200182	1026	36	Outpatient	NR	PCR+
Zigman Suchsland, 202183	739	NR	Population based	A+B	PCR
Zimmerman, 201684	4173	34.1	Outpatient primary care	A+B	PCR

.ED, emergency department; ILIinfluenza-like illnessNR, not reported

Study quality

Regarding assessment of study quality, 45 studies were judged to be at low risk of bias, 19 at moderate risk of bias and 3 at high risk of bias. The three studies at high risk of bias were downgraded due to failure to clearly have recruited consecutive patients, not accounting for all recruited patients,^{13 14} and use of inappropriate exclusion criteria.¹³ The full quality assessment is in online supplemental appendix table 2.

Individual signs and symptoms and their combinations

The accuracy of individual signs and symptoms and simple combinations of symptoms with a Youden index of 0.10 or greater is shown in table 3 (the full summary table including stratification by age can be found in tables3 4 and complete study-level data can be found in online supplemental appendix tables 6 and 7). We show 95% CIs for each measure of accuracy in the tables but have not reproduced them in the text for clarity. While many studies reported the accuracy of various combinations of symptoms, only the combinations of ‘cough+fever’ and ‘fever+sore throat’ were reported by at least five studies. The signs and symptoms with a Youden index greater than 0.15 were any subjective or measured fever (LR+ 1.7, LR− 0.30, Youden 0.32), the overall clinical impression (LR+ 2.1, LR− 0.63, Youden 0.28), coryza (LR+ 1.6, LR− 0.52, Youden 0.25), the combination of ‘cough+fever’ (LR+ 1.5, LR− 0.53, Youden 0.25), measured fever (LR+ 1.7, LR− 0.62, Youden 0.25), subjective fever (LR+ 1.4, LR− 0.49, Youden 0.22) and chills (LR+ 1.4, LR− 0.71, Youden 0.17). In addition to cough+fever, the combinations of ‘cough+fever+myalgias’ (LR+ 2.0, LR− 0.56, Youden 0.31, AUC 0.70) and ‘cough+fever+ sore throat’ (LR+1.9, LR− 0.63, Youden 0.26, AUC 0.69) had fair accuracy but were only reported in three studies (online supplemental appendix table 2).

Table 3. Accuracy of signs and symptoms, sorted by Youden Index and then by AUC.

Sign or symptom and age group	# of Studies	Sensitivity	Specificity	LR+ (95% CI)	LR− (95% CI)	AUC	Youden index
Fever (subjective or measured)	13	0.87 (0.79 to 0.93)	0.45 (0.24 to 0.69)	1.7 (1.2 to 2.7)	0.30 (0.17 to 0.50)	0.82	0.32
Overall impression	5	0.54 (0.35 to 0.71)	0.74 (0.57 to 0.86)	2.1 (1.6 to 2.7)	0.63 (0.50 to 0.75)	0.68	0.28
Coryza	8	0.76 (0.68 to 0.82)	0.49 (0.27 to 0.72)	1.6 (1.1 to 2.5)	0.52 (0.35 to 0.79)	0.73	0.25
Cough and fever	15	0.74 (0.57 to 0.85)	0.51 (0.29 to 0.72)	1.5 (1.2 to 2.2)	0.53 (0.42 to 0.66)	0.69	0.25
Fever (measured)	25	0.60 (0.50 to 0.70)	0.65 (0.54 to 0.76)	1.7 (1.5 to 2.1)	0.62 (0.55 to 0.68)	0.66	0.25
Fever (subjective)	27	0.81 (0.73 to 0.86)	0.41 (0.29 to 0.55)	1.4 (1.2 to 1.6)	0.49 (0.39 to 0.56)	0.71	0.22
Chills	29	0.61 (0.50 to 0.70)	0.56 (0.44 to 0.67)	1.4 (1.2 to 1.6)	0.71 (0.63 to 0.78)	0.61	0.17
Cough	72	0.92 (0.90 to 0.94)	0.22 (0.17 to 0.29)	1.2 (1.1 to 1.3)	0.36 (0.30 to 0.43)	0.77	0.14
Nasal congestion	13	0.69 (0.56 to 0.79)	0.45 (0.31 to 0.60)	1.3 (1.1 to 1.5)	0.71 (0.55 to 0.90)	0.60	0.14
Acute onset	17	0.52 (0.36 to 0.62)	0.60 (0.48 to 0.70)	1.3 (1.03 to 1.6)	0.81 (0.64 to 0.97)	0.58	0.12
Fatigue	38	0.69 (0.56 to 0.79)	0.42 (0.30 to 0.55)	1.2 (1.1 to 1.3)	0.74 (0.66 to 0.82)	0.57	0.11
Myalgia	59	0.59 (0.52 to 0.65)	0.52 (0.44 to 0.60)	1.2 (1.1 to 1.3)	0.79 (0.73 to 0.85)	0.57	0.11
Loss of appetite	10	0.51 (0.34 to 0.68)	0.60 (0.43 to 0.75)	1.3 (1.2 to 1.4)	0.83 (0.74 to 0.89)	0.57	0.11
Headache	55	0.65 (0.58 to 0.71)	0.45 (0.37 to 0.53)	1.2 (1.1 to 1.3)	0.78 (0.73 to 0.83)	0.57	0.10
Rhinorrhoea	39	0.70 (0.63 to 0.76)	0.40 (0.32 to 0.48)	1.2 (1.1 to 1.3)	0.75 (0.68 to 0.84)	0.57	0.10
Joint pain	11	0.29 (0.19 to 0.41)	0.78 (0.66 to 0.86)	1.3 (1.15 to 1.4)	0.92 (0.87 to 0.96)	0.54	0.07
Weakness	5	0.84 (0.74 to 0.90)	0.22 (0.11 to 0.37)	1.1 (1.01 to 1.2)	0.76 (0.65 to 0.92)	0.62	0.06
Chest pain	10	0.25 (0.13 to 0.43)	0.80 (0.65 to 0.89)	1.2 (1.01 to 1.5)	0.94 (0.85 to 0.99)	0.55	0.05
Sore throat	62	0.57 (0.51 to 0.63)	0.48 (0.41 to 0.55)	1.1 (1.0 to 1.2)	0.90 (0.84 to 0.96)	0.53	0.05
Rash	6	0.04 (0.01 to 0.09)	0.94 (0.86 to 0.97)	0.58 (0.50 to 0.66)	1.03 (1.01 to 1.07)	0.42	−0.02
Diarrhoea	29	0.14 (0.11 to 0.17)	0.75 (0.66 to 0.82)	0.57 (0.40 to 0.79)	1.2 (1.1 to 1.3)	0.25	−0.11
Vaccinated	11	0.27 (0.16 to 0.43)	0.62 (0.48 to 0.74)	0.71 (0.57 to 0.86)	1.17 (1.1 to 1.3)	0.44	−0.11

Only signs, symptoms and combinations with an LR significantly different from 1.0 are included; the full data stratified by age can be found in online supplemental appendix D, table A1. CIs are reported for signs or symptoms reported by five or more studies; otherwise ranges or a single study estimate are given.

AUC, area under the curve; DOR, diagnostic ORLRlikelihood ratio

Table 4. Accuracy of case definitions for the diagnosis of influenza.

Case definition and age group	Studies	Sensitivity	Specificity	LR+ (95% CI)	LR− (95% CI)	AUC	Youden index
WHO ILI (2018)
All studies	7	0.66 (0.48 to 0.81)	0.64 (0.39 to 0.83)	1.91 (1.30 to3.00)	0.54 (0.43 to0.66)	0.70	0.30
Adult	3	0.36 to 0.47	0.76 to 0.97
Infant, child or adult	4	0.70 to 0.90	0.21 to 0.52
ECDC FARI
All studies	7	0.70 (0.56 to 0.81)	0.59 (0.38 to 0.77)	1.76 (1.26 to2.59)	0.52 (0.41 to0.65)	0.70	0.29
Infant	2	0.64 to 0.77	0.27 to 0.44
Child or adult	4	0.45 to 0.86	0.37 to 0.84
Adult	1	0.50	0.90
CDC ILI
All studies	23	0.64 (0.49 to 0.77)	0.61 (0.40 to 0.78)	1.68 (1.27 to2.35)	0.60 (0.50 to0.70)	0.66	0.25
Infant	2	0.68 to 0.86	0.05 to 0.43
Child or adult	4	0.45 to 0.97	0.03 to 0.85
Adult	9	0.40 (0.21 to 0.61)	0.87 (0.67 to 0.95)	3.13 (1.71 to5.57)	0.70 (0.54 to0.84)	0.68	0.27
Infant, child or adult	7	0.78 (0.56 to 0.90)	0.39 (0.19 to 0.64)	1.30 (1.08 to1.75)	0.58 (0.37 to0.82)	0.63	0.17
WHO SARI (2011)
All studies	11	0.29 (0.13 to 0.53)	0.74 (0.49 to 0.90)	1.13 (0.89 to 1.44)	0.96 (0.87 to 1.06)	0.50	0.03
Infant	6	0.42 (0.16 to 0.73)	0.51 (0.19 to 0.82)	0.87 (0.75 to 1.01)	1.17 (0.99 to 1.48)	0.44	−0.07
Child or adult	5	0.19 (0.05 to 0.51)	0.87 (0.64 to 0.96)	1.43 (1.19 to1.70)	0.92 (0.75 to0.99)	0.60	0.06

Only those definitions evaluated in five or more studies are shown here; see online supplemental appendix D table A2 for the full set of studies.

Bold face indicates a Likelihood Ratio significantly larger or smaller than 1.0.

ARI, acute respiratory infection; CDC, Centers for Disease Control and Prevention; ECDC, European CDC; FARI, febrile acute respiratory infectionILI, influenza like illness; SARI, severe acute respiratory infection

Summary ROC curves for the most accurate signs and symptoms stratified by age group are shown in figure 2A–F. Subjective or measured ‘fever+coryza’ (figure 2A,C) are both fairly uniformly sensitive, but with widely varying specificity. The overall impression, ‘cough+fever’, measured fever and subjective fever (figure 2B, D–F) all show evidence of threshold effects, where sensitivity increases as specificity decreases. This is typically caused by different implicit or explicit cut-offs for defining an abnormal test.

Figure 2. (A–F) Summary receiver operating characteristic curves for the six most accurate individual signs and symptoms stratified by age group.

Symptoms significantly but more weakly associated with influenza (Youden 0.10–0.14) included cough (LR+ 1.2, LR− 0.36, Youden 0.14), nasal congestion (LR+ 1.3, LR− 0.71, Youden 0.14), acute onset (LR+ 1.3, LR− 0.81, Youden 0.12), fatigue (LR+ 1.2, LR− 0.74, Youden 0.11, myalgias (LR+ 1.2, LR− 0.79, Youden 0.11), loss of appetite (LR+ 1.3, LR− 0.83, Youden 0.11), headache (LR+ 1.2, LR− 0.78, Youden 0.10) and rhinorrhoea (LR+ 1.2, LR− 0.75, Youden 0.10).

The presence of joint pain, weakness, chest pain and sore throat were significantly but only very weakly associated with influenza (online supplemental appendix 3, Youden index 0.05−0.09). Abdominal pain, cough with sputum, wheeze, seizures, ear pain, confusion, conjunctivitis, sinus symptoms, impaired feeding, digestive symptoms, sweats or diaphoresis, dyspnoea, tachypnoea, nausea, vomiting and ‘nausea or vomiting’ had no significant association with the diagnosis of influenza based on non-significant LRs and Youden indices near 0.0. The presence of rash, history of vaccination and diarrhoea made influenza less likely (Youden index range from −0.02 to −0.11).

The Youden index generally paralleled the AUC, with the exception of any cough (AUC 0.79, Youden 0.14). Only the overall clinical impression had a positive LR (+LR) greater than 2.0; it was also the most specific finding (0.74). Cough was the most sensitive finding (92%). Findings with the lowest negative LR and therefore best at ruling out influenza when absent were absence of cough (LR− 0.36 overall and 0.28 in adults) and absence of any fever (LR− 0.30).

Relatively few symptoms had sufficient data to calculate summary estimates of accuracy by age group. For most other signs and symptoms, the positive and negative LRs were generally similar regardless of age group or groups. Fatigue and myalgias were less sensitive and overall less accurate for the diagnosis of influenza in infants than in older populations.

Case definitions of influenza-like illness and ARI

The accuracy of clinical case definitions stratified by age group is shown in table 4. The WHO ILI 2018 (LR+ 1.91, LR− 0.54, Youden 0.30) and European CDC (ECDC) febrile ARI (FARI) case definitions (LR+ 1.76, LR− 0.52, Youden 0.29) were most accurate overall. Slightly less accurate was the CDC ILI (23 studies, LR+ 1.68, LR− 0.60, Youden 0.25). The WHO SARI 2011 case definition had poor accuracy for diagnosis of influenza (11 studies, LR+ 1.13, LR− 0.96, AUC 0.50, Youden 0.03). The full table includes other definitions that were reported by four or fewer studies and are presented in online supplemental appendix tables 5 and 8 for study-level data.

Summary ROC curves for the four best-studied case definitions stratified by age group are shown in figure 3A–D. The European CDC and WHO case definitions for ILI had modest sensitivity and specificity. The WHO ILI case definition had low sensitivity and good specificity in studies done in adults, while studies that included infants and children as well as adults had better sensitivity but worse specificity. The CDC definition of influenza-like illness had a clear threshold effect, which may be due to implicit differences in determining when a cough, sore throat or subjective fever was present between studies. As with the WHO ILI, studies in adults tended to have lower sensitivity and better specificity. The WHO SARI 2011 case definition was unhelpful for diagnosing influenza. See online supplemental appendix table A2 for the complete data.

Figure 3. (A–D) Summary receiver operating characteristic curves for the four best-studied case definitions stratified by age group. CDC, Centers for Disease Control and Prevention; ECDC, European CDC; FARI, febrile acute respiratory infection; ILI, influenza-like illness; SARI, severe acute respiratory infection.

Discussion

We have performed a comprehensive systematic review of signs, symptoms, their combinations and case definitions for the diagnosis of influenza. We found that individual signs and symptoms are of very limited value for distinguishing influenza from other causes of respiratory symptoms. The most accurate individual signs and symptoms are any ‘subjective or measured fever’ (LR+1.7, LR− 0.30), the overall clinical impression (LR+ 2.1, LR− 0.63), coryza (LR+1.6, LR− 0.52) and the combination of cough+fever (LR+1.5, LR− 0.53). Absence of any fever (LR− 0.30) and absence of cough (LR− 0.36 overall, LR− 0.28 in adults) were the individual signs and symptoms best at ruling out influenza when absent. The most sensitive individual findings were fever (0.87) and cough (0.92), while the most specific single finding was the overall clinical impression (0.74).

There was considerable heterogeneity with regard to the accuracy of signs and symptoms as shown in figure 2A–F. A threshold effect, with sensitivity and specificity varying inversely due to different explicit or implicit thresholds for a positive test were especially apparent for overall impression, ‘cough+fever’, measured fever and subjective fever. There were also some differences by age group: any fever appeared to be less sensitive in adults than in younger groups, while fatigue was much more sensitive in adults than in infants (0.82 vs 0.27). Not surprisingly, symptoms relying on patient report such as fatigue and myalgia were unhelpful in infants. There was good consistency across age groups for cough as a predictor of influenza. Symptoms previously found to be independent predictors of influenza such as acute onset (LR+1.3, LR− 0.81) and myalgia¹⁵ (LR+ 1.2, LR− 0.79) had only modest accuracy in our systematic review.

Among combinations of symptoms, ‘cough+fever’ was best studied but had only modest accuracy (LR+1.5, LR− 0.53). In a separate systematic review, we identified 14 clinical prediction rules (CPRs) that used combinations of signs and symptoms to predict the likelihood of influenza.¹⁶ However, none has been successfully validated in an independent population. Future studies to develop and validate CPRs should include the most accurate individual signs and symptoms identified by our analysis, including variables not previously included such as coryza and the overall clinical impression, as well as signs and symptoms found to decrease the likelihood of influenza such as vaccination and gastrointestinal symptoms. Different CPRs are likely needed for children and adults.

The WHO case definition for ILI was the most accurate overall (8 studies, LR+ 1.91, LR− 0.54, AUC 0.70), while for adults only the CDC ILI was most accurate (9 studies, LR+ 3.1, LR− 0.70, AUC 0.68).

In a population with a 25% prevalence of influenza, an adult meeting the CDC case definition would have a 51% likelihood of influenza. However, not meeting the case definition only reduces that likelihood to 19%. The 2018 version of the WHO ILI case definition and the European CDC FARI case definition had similar overall accuracy to the CDC case definition, but were less well studied, especially in subgroups by age. Thus, other than the ability of the CDC ILI case definition to help rule out influenza in adults when criteria are met, case definitions do not appear to be helpful to clinicians for the diagnosis of influenza.

The assessment of signs and symptoms is inherently subjective and may be influenced by clinician expectations around the likelihood of influenza, which is in turn driven by prevalence in the community at that time. The studies in this systematic review were done during times of year when influenza was circulating. Incorporating the prevalence of a pathogen using surveillance data is a promising way to enhance the accuracy of diagnostic risk scores¹⁷ and could be applied to influenza. The assessment of signs and symptoms can also be influenced by implicit differences in what is classified as coryza or sore throat or cough. For example, any cough or only moderate or severe cough? This is captured by the threshold effects seen in many of the ROC curves in figures2 3.

Strengths and limitations

Our analysis had several notable strengths. The search was comprehensive and identified 67 studies, most studies were judged to be at low risk of bias, and the analysis used a bivariate approach with stratification by age group. All studies gathered data prospectively and used high-quality reference standards. Also, all studies but one gathered data prior to the current COVID-19 pandemic. A limitation was that for many of the signs and symptoms, there were relatively few studies reporting accuracy separately for infants, children and adults.

Implications for clinicians, researchers, policy-makers

For clinicians, it is clear that using either individual signs and symptoms or existing combinations of clinical features has very limited accuracy for identifying adults or children with influenza, and likely in differentiating influenza from COVID-19 and other respiratory pathogens.

Diagnostic tests to accurately identify patients who have influenza, perhaps used in conjunction with CPRs, have the potential to identify patients who might benefit from antiviral drugs and reduce inappropriate antibiotic use.¹⁸ The development of more accurate molecular tests combining detection of influenza virus with SARS-CoV-2 and possibly other pathogens such as respiratory syncitial virus (RSV) in children is promising.¹⁹ While the cost of point-of-care molecular tests remains high, more accurate diagnosis of highly contagious infections has the added benefit of allowing the institution of isolation and mask-wearing to reduce transmission to others. Dissemination of cheap rapid antigen tests to low-resource settings is also needed.

With increasing availability of point-of-care tests, including in low-resource countries, a pragmatic strategy for clinical practice would be initially using a combination of signs and symptoms that is highly sensitive with adequate specificity (to miss as few cases as possible) and then using a point-of-care test to confirm the diagnosis. Fever and cough (individually) both have high sensitivity, with fever having higher specificity. We did not find a combination of fever and any respiratory symptom that would be worth exploring as an initial filter.

For researchers, more work is needed to validate existing CPRs, as well as to develop and validate new ones. We recently published a systematic review of CPRs for the diagnosis of influenza and identified a number with very good accuracy that were able to risk stratify patients into low, moderate and high-risk groups. This kind of risk stratification better fits the clinical practice of rule out, need more information, and rule in for low, moderate and high-risk groups. However, none of the CPRs were successfully prospectively validated. Future work is needed to develop and validate new or existing CPRs for influenza.

For example, an individual patient meta-analysis using combined datasets with similar inclusion criteria from different populations could be used to develop CPRs in one set of populations and validate them in others. Integrating the overall impression and creating separate rules for adults and children will also be important. CPRs using only symptoms can facilitate evidence-based triage decisions and telemedicine care, perhaps in conjunction with the availability of home rapid antigen tests for influenza. We recently studied the impact of such a home test on physician decision-making and found that the availability of such a test could have a major impact on the need for in-person consultations, potentially reducing overall cost and increasing convenience (manuscript in review). It could also help avoid transmission of infection to other patients in the outpatient practice and to staff.

For policy-makers, it is important to understand the poor accuracy of existing case definitions for influenza-like illness. This limits their ability to perform accurate disease surveillance. In fact, we found no studies reporting the accuracy of a case definition in a population of children only, yet clinicians and policy-makers in low-resource settings often rely on these case definitions in the absence of diagnostic testing to make clinical and policy decisions about the care of children with respiratory infections. Better case definitions are needed with evaluation in children and/or more widespread dissemination of point-of-care tests.

The funder was not involved in the design of the study, does not have any ownership over the management and conduct of the study, the data, or the rights to publish.

Data availability statement

Data are available on reasonable request.