Sample characteristics
A total of 1085 influenza virus infections were diagnosed in adults during the development stage of this study, 482 in the influenza season of 2016-2017, and 603 in the 2017-2018 season. After inclusion and exclusion criteria were applied, 707 patients remained (figure 1). Influenza A was the most frequent subtype (561/707, 79.3%), followed by influenza B (145/707, 20.5%), while only one patient (0.01%) had influenza A and B coinfection.
Three hundred ninety-four infections were diagnosed in the validation stage, with 285 patients finally included in the external validation cohort.
Patients’ characteristics are reported in table 1. In the derivation cohort, median age was 79 (65-85) years, 52.6% of patients were male, and comorbidity as moderate (median Charlson Comorbidity Index 2, 1-5). The median of oxygen saturation (SpO2) on room air at admission was 92% (88-95) and median respiratory rate was 18 (16-24) breaths per minute. Radiological abnormalities were present in 40.5% of patients, and the median lymphocyte count was 700 cells/µL (400-1000). Forty-eight patients (6.8%) developed RF during hospitalization. From all patients, 4.2% were admitted to an intensive care unit (ICU) irrespective of the reason, 2.5% received NIPPV, and 2.7% received IMV. 43.8% of the patients with RF needed ICU admission and 62.5% received ventilatory support. Thirty-four patients (4.8%) died during hospitalization.
In the external validation cohort, all influenza isolates corresponded to influenza A. Median duration of symptoms before hospital admission was shorter (3 (1-4) days vs 3 (2-6) days, p<0.001), and more patients were vaccinated (54.5% vs 46.0%, p=0.017) in the validation cohort. Radiological findings were similar (44.9% vs 40.5%, p>0.20), but fewer secondary pneumonia episodes were observed in the second cohort (9.5% vs 19.1%, p<0.001). Thirty-three patients developed RF in this cohort (11.6% vs 6.8%, p=0.013). No differences were observed in mortality or ICU admissions, but more patients received NIPPV (6.3% vs 2.5%, p=0.004) in the validation cohort.
Table 1
General characteristics of the samples.
Variable, n (%)
|
Total
(N=992)
|
Derivation cohort
(n=707)
|
Validation cohort
(n=285)
|
p
|
Influenza subtype
- A
- B
- Both
|
846 (85.3%)
145 (14.5%)
1 (0.1%)
|
561 (79,3%)
145 (20.5%)
1 (0.1%)
|
285 (100.0%)
0 (0.0%)
0 (0.0%)
|
<0.001
|
Age (median, IQR)
|
79 (66-85)
|
79 (65-85)
|
79 (67-85)
|
>0.20
|
Sex (male)
|
522 (52.6%)
|
372 (52.6%)
|
150 (52.6%)
|
>0.20
|
Days since symptoms (median, IQR)
|
3 (2-5)
|
3 (2-6)
|
3 (1-4)
|
<0.001
|
Symptoms lasting >72 hours
|
427 (43.6%)
|
327 (47.1%)
|
100 (35.1%)
|
0.001
|
Influenza vaccine
|
480 (48.4%)
|
325 (46.0%)
|
155 (54.4%)
|
0.017
|
Hypertension
|
672 (67.7%)
|
484 (68.5%)
|
188 (66.0%)
|
>0.20
|
Pregnancy
|
9 (0.9%)
|
7 (1.0%)
|
2 (0.7%)
|
>0.20
|
Current smokers
|
126 (12.8%)
|
91 (12.9%)
|
35 (12.7%)
|
>0.20
|
Obesity
|
196 (45.6%)
|
108 (46.0%)
|
88 (45.1%)
|
>0.20
|
Asthma
|
90 (9.1%)
|
73 (10.3%)
|
17 (6.0%)
|
0.03
|
COPD
|
242 (24.4%)
|
172 (24.3%)
|
70 (24.6%)
|
>0.20
|
Diabetes mellitus
|
276 (27.8%)
|
206 (29.1%)
|
70 (24.6%)
|
0.15
|
Cardiovascular disease
|
348 (35.1%)
|
240 (33.9%)
|
108 (37.9%)
|
>0.20
|
Chronic kindney disease
|
176 (17.7%)
|
127 (18.0%)
|
49 (17.2%)
|
>0.20
|
Liver disease
|
53 (5.3%)
|
36 (5.1%)
|
17 (6.0%)
|
>0.20
|
Neurologic disorder
|
170 (17.1%)
|
188 (16.7%)
|
52 (18.2%)
|
>0.20
|
Immunosupression (*)
|
177 (17.8%)
|
132 (18.7%)
|
45 (15.8%)
|
>0.20
|
CCI (median. IQR)
|
2 (1-4)
|
2 (1-5)
|
2 (1-3)
|
<0.001
|
CCI >2 points
|
620 (62.8%)
|
458 (65.2%)
|
162 (56.8%)
|
0.01
|
Respiratory rate (median, IQR)
|
20 (16-25)
|
18 (16-24)
|
22 (16-28)
|
<0.001
|
pO2 on room air (median, IQR)
|
56 (50-64)
|
57 (50-65)
|
54 (50-61)
|
0.004
|
SpO2 on room air (median, IQR)
SpO2 on room air <94%
|
91 (88-95)
627 (66.6%)
|
92 (88-95)
440 (66.4%)
|
91 (87-95)
187 (67.0%)
|
0.099
>0.20
|
Radiological abnormalities
|
396 (41.7%)
|
274 (40.5%)
|
122 (44.9%)
|
>0.20
|
Secondary pneumonia
|
161 (16.4%)
|
135 (19.1%)
|
26 (9.5%)
|
<0.001
|
Lymphocyte count (cells/µl) (median, IQR)
|
700 (400-1000)
|
700 (400-1000)
|
700 (400-1000)
|
>0.20
|
LDH (U/l) (median, IQR)
|
275 (224-335)
|
287 (246-348)
|
228 (188-292)
|
<0.001
|
Respiratory failure
|
81 (8.2%)
|
48 (6.8%)
|
33 (11.6%)
|
0.01
|
Admission to ICU
|
49 (4.9%)
|
30 (4.2%)
|
19 (6.7%)
|
0.11
|
Mechanical Ventilation
- NIPPV
- IMV
|
36 (3.6%)
22 (2.2%)
|
18 (2.5%)
19 (2.7%)
|
18 (6.3%)
3 (1.1%)
|
0.004
0.11
|
Mortality
- In hospital (overall)
- 30-day
|
47 (4.8%)
25 (2.6%)
|
33 (4.7%)
18 (2.5%)
|
14 (4.9%)
7 (2.6%)
|
>0.20
>0.20
|
Evolution after discharge
- Recovery without sequelae
- Recovery with sequelae
|
794 (80.3%)
148 (15.0%)
|
592 (84.1%)
79 (11.2%)
|
202 (70.9%)
69 (24.2%)
|
<0.001
|
Length of hospital stay (days. median. IQR)
|
7 (5-13)
|
7 (5-12)
|
8 (5-16)
|
0.013
|
COPD: chronic obstructive pulmonary disease; CCI: Charlson Comorbidity Index; pO2: partial pressure of oxygen; SpO2: oxygen saturation; LDH: lactacte dehydrogenase; ICU: Intesive Care Unit; NIPPV: non-invasive positive pressure ventilation; IMV: invasive mechanical ventilation.
(*)Immunosuppression was defined as the presence of any of: active cancer, autoimmune disease, solid organ or hematological transplantation, HIV infection, treatment with immunosuppressive drugs or active chemotherap
Predictive model
Variables included into the multivariate logistic regression model and those remaining after backwards stepwise elimination are reported in table 2. Lymphocytes at admission were kept in on grounds of prognostic implications in the extant literature and a statistical significance of p<0.10.
Missing value imputation
The only variable with more than 10% missing values from those included in the models was respiratory rate (245 missing). Missing respiratory rate values were imputed five times using multiple imputation, resulting in six sets of data. Statistical analysis and performance assessment was conducted once in each set.
Derivation of clinical prediction rules
A formula for RF prediction was created using coefficients of the final logistic regression model but, due to the predicted event being infrequent, low sensitivity was achieved. Since the aim of the study was to develop a scale of risk categories, recursive partitioning was used to ascribe risk scores. Decision trees were performed separately for each
Table 2
Univariate and multivariate logistic regression analysis.
VARIABLE
|
Univariate OR (CI 95%)
|
p
|
Multivariate OR (CI 95%)
|
p
|
FM OR (CI 95%)
|
p
|
Age*
|
0.980 (0.964-0.997)
|
0.023
|
0.977 (0.955-0.999)
|
0.037
|
|
|
Sex (male)
|
1.339 (0.744-2.410)
|
0.331
|
|
|
|
|
Influenza vaccine
|
0.623 (0.338-1.148)
|
0.129
|
|
|
|
|
Hypertension
|
0.753 (0.410-1.381)
|
0.359
|
|
|
|
|
Current smokers
|
1.624 (0.759-3.475)
|
0.212
|
|
|
|
|
Obesity
|
0.610 (0.254-1.467)
|
0.270
|
|
|
|
|
COPD
|
1.783 (0.961-3.309)
|
0.067
|
1.980 (0.954-4.109)
|
0.067
|
|
|
Diabetes mellitus
|
0.622 (0.304-1.272)
|
0.193
|
|
|
|
|
Immunossupresion
|
0.604 (0.251-1.453)
|
0.260
|
0.671 (0.265-1.704)
|
0.671
|
|
|
CCI
|
1.030 (0.922-1.151)
|
0.599
|
|
|
|
|
Respiratory rate*
|
1.114 (1.090-1.200)
|
<0.001
|
1.112 (1.055-1.172)
|
<0.001
|
1.110 (1.056-1.168)
|
0.000
|
SpO2 on room air*
|
0.927 (0.895-0.960)
|
<0.001
|
|
|
|
|
Radiological abnormalities
|
2.093 (1.149-3.814)
|
0.016
|
2.826 (1.333-5.991)
|
0.007
|
2.715 (1.341-5.498)
|
0.006
|
Secondary pneumonia
|
3.067 (1.663-5.657)
|
<0.001
|
|
|
|
|
Platelets (cells/µl)*
|
1.002 (0.999-1.005)
|
0.224
|
|
|
|
|
Lymphocytes (cells/µl)*
|
1.008 (0.995-1.021)
|
0.217
|
1.000 (1.000-1.000)
|
0.070
|
1.000 (1.000-1.000)
|
0.094
|
CRP (mg/dl)*
|
1.020 (0.992-1.049)
|
0.160
|
1.002 (0.968-1.037)
|
0.929
|
|
|
LDH (U/l)*
|
1.004 (1.002-1.005)
|
<0.001
|
1.003 (1.001-1.005)
|
0.002
|
1.003 (1.002-1.005)
|
0.000
|
OR: Odds ratio. CI: confidence Interval. FM: final multivariate model after backwards elimination. COPD: Chronic Obstructive Pulmonary Disease. CCI: Charlson Comorbidity Index. SpO2: oxygen saturation. CRP: C-reactive protein. LDH: lactate dehydrogenase. *Odds Ratio per increment of 1 unit.
variable since no interaction or collinearity was found. Scores were assigned according to the RF proportion in the different intervals of continuous variables: zero points assigned to intervals with less than 5% RF, one point assigned to intervals with 5-10% RF, and two points assigned to intervals with more than 10% RF. Two points were assigned to patients with radiological abnormalities considering and Odds Ratio of 2 in the multivariate logistic regression. The score was manually fine-tuned using ROC curves, and simplified to a maximum possible score of 7 points, and scores were grouped into four categories (figure 2).
The final risk score is computed by assigning 0-2 points to the lymphocyte count at admission, 0-2 points to the lactate-dehydrogenase (LDH) levels at admission, 0-2 points to the respiratory rate at admission, and 1 point if radiological abnormalities are present. This score is then converted to risk categories, with 0 points corresponding to category A, 1-2 points corresponding to category B, 3-4 points to category C, and 5 or more points to category D. RF risk was computed for each category using 1000 sample simulation bootstrapping (table 3). Precision analysis is reported in Supplementary material 1.
Table 3
Risk score and risk categories.
VARIABLE
|
SCORE
|
Lymphocytes at admission
≥600 cells/µl
400-600 cells/µl
≤400 cells/µl
|
0 points
1 points
2 points
|
Respiratory rate
≤20 brpm
21-28 brpm
>28 brpm
|
0 points
1 points
2 points
|
LDH at admission
≤280 U/l
280-400 U/l
>400 U/l
|
0 points
1 points
2 points
|
Radiological alteration
No
Yes
|
0 points
1 points
|
RISK CATEGORY
|
Respiratory failure proportion (CI 95%)
|
A: 0 points
|
0.0% (0.0%-0.0%)
|
B: 1-2 points
|
3.5% (1.0%-6.6%)
|
C: 3-4 points
|
10.0% (5.0%-15.8%)
|
D: 5 or more points
|
45.2% (29.0%-61.3%)
|
brpm: breath per minute; CI: Confidence Interval. |
Validation of the clinical prediction rule
Internal validation was performed using five-fold cross-validation. The clinical prediction rule maintained its discrimination capacity in the five subsets within the derivation cohort (Supplementary material 2).
Risk scores and categories were computed for patients in the external validation cohort. ROC curves for both cohorts are shown in figure 3. Area under the ROC curve was 0.796 (0.714-0.877) in the derivation cohort and 0.773 (0.687-0.859) in the validation cohort (p-value <0.001 in both cases). Convenient classification power was observed in both cohorts, with a conclusive Chi-squared test for trend for risk categories and RF development (p-value <0.05 in all cases). An adequate fit was observed between predicted and observed RF proportions in the four categories (Fisher’s exact test for goodness-of-fit p-value=0.43; Chi-squared test for goodness-of-fit with merging of 0 predicted event categories p-value=0.42).