Developing laboratory indicators to predict the probability of in-hospital progression to severe disease in elderly patients with influenza A virus infection

Abstract

Background

The progression of respiratory viral infections to severe disease, and the associated high mortality in elderly patients, pose a significant clinical challenge. Currently, effective warning tools specific for predicting severe influenza are lacking.

Objectives

To preliminarily develop a laboratory‑based prediction model using routine clinical laboratory markers to estimate the risk of in-hospital progression to severe disease in elderly patients with influenza A virus infection.

Methods

This retrospective cohort study was conducted at Shanghai Fourth People’s Hospital, affiliated with Tongji University, from March 2023 to February 2025. Data were extracted from the hospital’s electronic medical record system. A total of 198 elderly patients (aged > 60 years) with laboratory-confirmed influenza A virus infection were enrolled; those co-infected with other respiratory pathogens were excluded. All patients had complete demographic and outcome data, and laboratory biomarkers were measured within 24 h of admission. The primary outcome was progression to severe influenza, defined according to the Chinese “Diagnosis and Treatment Protocol for Influenza” (2025 Edition). Patients were randomly split into a training set and a validation set in a 7:3 ratio. Elastic net regression was employed to select candidate variables for constructing a prognostic model of disease severity. Model performance was evaluated based on discrimination and calibration.

Results

Of the 198 elderly hospitalized patients with influenza A virus infection, 50 developed severe disease, while 148 remained non-severe. There were no significant differences in age or sex between the two groups. However, 16 clinical laboratory markers, including procalcitonin, interleukin-6, and albumin, showed significant differences. Using a machine learning-based variable selection procedure, we identified white blood cell count, procalcitonin, interleukin-6, and myoglobin and incorporated them into the prediction model. The model showed discriminatory performance in both the training set (C-index: 0.851, 95% CI: 0.797–0.905) and the validation set (C-index: 0.819, 95% CI: 0.709–0.929). Calibration analysis suggested potentially acceptable calibration at day 14.

Conclusions

We developed and preliminarily evaluated a laboratory‑based prognostic model to predict the risk of in-hospital progression to severe disease in elderly patients with influenza A virus infection. The model showed moderate discriminative ability in this single‑center cohort. Given the modest sample size and limited number of events, all results should be interpreted with caution and considered preliminary and exploratory. It will be important to conduct external validation in larger, independent cohorts.

Trial registration

The study protocol was retrospectively registered with the Chinese Clinical Trial Registry (ChiCTR2500112511) on November 14, 2025, after the study had been conducted.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12879-026-14007-w.

Introduction

The widespread implementation of non-pharmaceutical interventions (NPIs) during the COVID-19 pandemic significantly changed the epidemiological patterns of respiratory viruses [1]. As NPIs are gradually relaxed, several respiratory pathogens, including influenza A virus, influenza B virus, and respiratory syncytial virus, have demonstrated unusual non-seasonal and asynchronous resurgences, leading to repeated seasonal and sporadic outbreaks [2, 3]. The influenza A virus is a particularly concerning pathogen due to its epidemic and pandemic potential, largely driven by two key mechanisms of variation: antigenic drift and antigenic shift. Antigenic drift refers to the gradual accumulation of point mutations in viral genes. These mutations result in minor changes to surface antigens, allowing the virus to evade existing host immunity. This mechanism underlies seasonal epidemics and necessitates annual vaccine updates. In contrast, antigenic shift, the fundamental driver of pandemics, occurs when gene segments reassort during co-infection with different viral strains. This process generates novel subtypes to which the population lacks broad immunity, potentially leading to global spread [4–6]. Due to its rapid and unpredictable mutation rate, the influenza A virus remains a significant pandemic threat [7].

Beyond the four major global pandemics caused by the influenza A virus, seasonal influenza also presents a considerable public health challenge [8]. The World Health Organization estimates that influenza causes roughly 3 to 5 million severe cases worldwide annually, leading to hundreds of thousands to millions of deaths related to the influenza A virus [9, 10]. Individuals aged 65 and above account for the largest proportion of influenza A virus-related deaths, making them the primary risk group [11]. Following influenza A virus infection, severe cases in this population often develop serious complications such as bacterial or fungal pneumonia, acute kidney injury, myocardial injury, sepsis-like syndrome, multiple organ failure, and disseminated intravascular coagulation [12, 13]. These complications pose significant challenges for clinical management and strain healthcare systems considerably.

Regarding prevention, annual influenza vaccination remains an effective strategy for reducing severe cases and mortality associated with influenza A virus infection. However, repeated vaccination can lead to reduced protection due to immune interference. This, combined with the high variability of influenza A, means that traditional prevention strategies relying solely on vaccination face inherent limitations [14–16]. In this context, prompt and accurate antiviral treatment following infection is of particular importance. Currently available clinical antiviral drugs for influenza A include Oseltamivir, Baloxavir marboxil, Zanamivir, Peramivir, and Favipiravir. However, their clinical efficacy varies substantially between individuals, offers limited symptomatic relief, and is highly dependent on timely administration [9, 17], particularly in high-risk groups such as patients with chronic diseases, the elderly, pregnant women, and young children. Considering the limitations of current drugs and the influenza A virus’s high variability and efficient transmission, vaccination and antiviral medications alone are insufficient to prevent severe outcomes effectively. Early diagnosis and detection are therefore crucial.

Currently, early warning tools for predicting severe progression of influenza A infection in elderly patients are lacking. While existing scoring systems like CURB-65, PSI, APACHE II, SOFA, and qSOFA are important for risk stratification in patients with community-acquired pneumonia and critically ill patients in intensive care units, their ability to identify influenza A-specific severe risk is limited. Furthermore, simple and reliable biomarkers for predicting severe influenza A progression are unavailable [18–20]. To address this gap, this exploratory study used readily available clinical biomarkers and multiple machine learning algorithms to develop a laboratory‑based predictive model for severe influenza A. The model’s incremental value beyond simple clinical information is not addressed here.

Materials and methods

Study population

This single-center, retrospective study was conducted at Shanghai Fourth People’s Hospital between March 2023 and February 2025. A total of 230 elderly patients (aged > 60 years) with laboratory-confirmed influenza A virus infection by real-time multiplex RT-PCR were initially identified. Among them, 12 patients were excluded due to missing clinical data. Additionally, 20 patients had already progressed to severe disease by the time of admission (due to delayed presentation) and were excluded, as our study aimed to estimate the risk of progression to severe disease after admission. Patients co-infected with other respiratory pathogens (e.g., influenza B virus, respiratory syncytial virus, or SARS-CoV-2) were also excluded. However, no such cases were identified in this cohort. After these exclusions, 198 patients who were non-severe at admission were included in the final analysis. These patients were divided into two groups: a severe group (n = 50) comprising patients who progressed to severe disease during hospitalization, and a non-severe group (n = 148) comprising those who did not. Severe disease was defined according to the Chinese “Diagnosis and Treatment Protocol for Influenza” (2025 Edition). Adults were classified as severe cases if they met any of the following criteria: (1) respiratory failure requiring mechanical ventilation; (2) shock; (3) acute necrotizing encephalopathy; or (4) concurrent organ failure requiring intensive care unit monitoring and treatment. All included patients had complete demographic and outcome data, and laboratory biomarkers were measured within 24 h of admission, although some biomarker values were missing. The detailed enrollment process is shown in Fig. 1.

Fig. 1.

Flowchart of patient selection

The study protocol adhered to the ethical principles outlined in the Declaration of Helsinki (1975) [21]. It received approval from the Ethics Committee of Shanghai Fourth People’s Hospital, affiliated with Tongji University (No. 2025079-001), and was retrospectively registered with the Chinese Clinical Trial Registry (ChiCTR2500112511) after data collection. The informed consent of all patients was waived for the retrospective study.

Predictors

We extracted demographic characteristics (age, gender, and medical history) and clinical data (laboratory biomarkers, treatments, and outcomes) for each admitted patient from the hospital’s electronic medical record database.

The primary variables include a total of 34 laboratory biomarkers: (1) Routine blood tests: white blood cell count (WBC, 10⁹/L), hemoglobin (Hb, g/L), lymphocyte percentage (LYM%), neutrophil percentage (NEUT%); (2) Inflammatory markers: C-reactive protein (CRP, mg/L), serum amyloid A (SAA, mg/L), procalcitonin (PCT, ng/mL), interleukin-6 (IL-6, pg/mL); (3) Renal function: serum creatinine (sCr, µmol/L), blood urea nitrogen (BUN, mmol/L); (4) Liver function: total protein (TP, g/L), albumin (ALB, g/L), total bilirubin (TBIL, µmol/L), alanine aminotransferase (ALT, U/L), aspartate aminotransferase (AST, U/L), alkaline phosphatase (ALP, U/L), gamma-glutamyl transferase (GGT, U/L); (5) Blood glucose: glucose (GLU, mmol/L); (6) Electrolytes: potassium ion (K⁺, mmol/L), sodium ion (Na⁺, mmol/L), chloride ion (Cl⁻, mmol/L); (7) Cardiac enzymes: creatine kinase-myocardial band (CK-MB, ng/L), myoglobin (MYO, ng/L), troponin T (TnT, ng/mL), N-terminal pro-B-type natriuretic peptide (NT-proBNP, pg/mL), B-type natriuretic peptide (BNP, pg/mL); (8) Coagulation factors: prothrombin time (PT, s), activated partial thromboplastin time (APTT, s), thrombin time (TT, s), fibrinogen (FIB, g/L), D-dimer (D-D, mg/L), fibrin(ogen) degradation products (FDP, µg/mL); (9) Arterial blood gas: pH, oxygen saturation (SaO₂, %).

Statistical methods

The primary outcome was progression to severe influenza, defined according to the criteria in the Chinese “Diagnosis and Treatment Protocol for Influenza” (2025 Edition). The time origin was defined as the date of hospital admission. The event time was defined as the first day a patient met any of these criteria. Patients who met the criteria and were subsequently discharged or died were considered to have experienced the event at the time the criteria were first met. Patients who were discharged without meeting the criteria for severe influenza were censored at the time of discharge. Because discharge may act as a competing event, we also performed a competing-risk analysis using the cumulative incidence function. Patients were followed from admission until in-hospital death or discharge.

The data were statistically processed and graphed using the R language. Prior to analysis, biomarkers with more than 5% missing data were excluded. Among the 34 laboratory biomarkers, 7 biomarkers were removed due to high missingness: FIB (25.8%), SAA (19.7%), NT-proBNP (43.4%), BNP (22.7%), pH (32.8%), SaO₂ (33.3%), and FDP (34.8%). The remaining 27 biomarkers, which were considered as the initial candidate variables, had missing rates ranging from 0% to 4.5%. 153 patients (77.3%) had complete data for all retained variables. Missingness was predominantly sporadic and confined to single variables. Patients were then randomly divided into a training set (70%) and a validation set (30%), and balance tests were conducted to confirm consistency between the two groups. For variables with less than 5% missing data, multiple imputation was performed separately in the training and validation sets using the R package mice. We generated five imputed datasets with five iterations each. Given the low missing rate (< 5%), we combined the imputed datasets by taking the mean of the imputed values across the five datasets for each missing entry and then performed variable selection.

In the training set, elastic net regression was used to select relevant variables from the 29 final candidate predictors, including the 27 retained biomarkers together with age and sex. This method simultaneously optimizes variable selection and coefficient shrinkage, effectively reducing the coefficients of irrelevant variables to zero. This makes it particularly well-suited for high-dimensional data with limited sample sizes. Furthermore, elastic net regression effectively addresses multicollinearity among predictor variables, a common issue with clinical indicators such as inflammatory markers, which are often highly correlated [22]. We then used Cox regression to obtain clinically interpretable hazard ratios and 95% confidence intervals for the variables selected by elastic net regression (i.e., those with non-zero coefficients) [23, 24]. To ensure the Cox model’s reliability, the proportional hazards assumption was tested. Based on the model, a nomogram was constructed to visualize risk prediction.

The model’s discriminative ability was evaluated using receiver operating characteristic (ROC) curves, area under the ROC curve (AUC), concordance index (C-index), and time-dependent AUCs in both the training and validation sets. Calibration curves assessed the consistency between predicted probabilities and actual incidence rates. Furthermore, we performed a bootstrap-based internal validation (1,000 resamples) to assess the stability of the model’s discriminative ability, while keeping the model coefficients fixed at those derived from the training set. The 95% confidence intervals for the time-dependent AUCs and the C-index were derived from the bootstrap distributions [25]. However, this bootstrap procedure did not repeat the full modeling pipeline (including missing data handling, elastic net variable selection, and Cox regression fitting). Therefore, the bootstrap analysis did not account for the uncertainty introduced by variable selection and model fitting.

Results

Baseline characteristics of elderly hospitalized patients with influenza A virus infection and event-time distribution

This study enrolled 198 eligible hospitalized patients. Of these, 50 patients progressed to severe disease (Severe group), while 148 patients did not (Non-severe group). In the Severe group, 26 patients were male (52.0%) and 24 were female (48.0%), with a median age of 78 years (IQR 68.5–88.0). The Non-severe group included 84 males (56.8%) and 64 females (43.2%), with a median age of 74 years (IQR 68.0-84.5). There were no significant differences in age or gender between the two groups. The most common underlying diseases among all patients were hypertension (57.1%) and heart disease (42.4%), followed by brain disease, diabetes, history of surgery, chronic obstructive pulmonary disease, liver disease, renal disease, non-lung malignancies, and neurological disorders. However, the Severe group had a significantly higher proportion of patients with comorbid brain disease (P = 0.049), heart disease (P = 0.016), liver disease (P = 0.026), and renal disease (P = 0.002) compared to the Non-severe group (Table 1).

Table 1.

The demographic characteristics of patients in the severe group and the non-severe group

Characteristic	Total (n = 198)	Severe (n = 50)	Non-severe (n = 148)	P-value
Gender, No. (%)				0.674
Male	110(55.6%)	26(52.0%)	84(56.8%)
Female	88(44.4%)	24(48.0%)	64(43.2%)
Age (years), median (IQR)	75.0 (68.0–87.0)	78.0 (68.5–88.0)	74.0 (68.0-84.5)	0.265
Comorbidity
Brain disease	74(37.4%)	25(50.0%)	49(33.1%)	0.049
Hypertension	113(57.1%)	32(64.0%)	81(54.7%)	0.327
Diabetes	71(35.9%)	19(38.0%)	52(35.1%)	0.846
History of surgery	79(39.9%)	20(40.0%)	59(39.9%)	1.000
Heart disease	84(42.4%)	29(58.0%)	55(37.2%)	0.016
COPD	22(11.1%)	9(18.0%)	13(8.8%)	0.125
Liver disease	27(13.6%)	12(24.0%)	15(10.1%)	0.026
Renal disease	36(18.2%)	17(34.0%)	19(12.8%)	0.002
Non-lung malignancies	39(19.7%)	13(26.0%)	26(17.6%)	0.275
Neurological disorders	28(14.1%)	11(22.0%)	17(11.5%)	0.107

COPD, Chronic obstructive pulmonary disease

In the study, 50 patients (25.3%) progressed to severe disease during hospitalization. The cumulative incidence of progression, estimated by the Kaplan-Meier method, was 30.1% (95% CI: 22.0-38.3%) at day 14. The median time to progression could not be estimated as the event rate did not reach 50%. The lower bound of the 95% confidence interval for the median was 17 days (Supplementary Figure S1). Among the 148 patients who did not progress to severe disease, 112 (75.7%) were discharged before day 14. To account for discharge as a competing event, we also performed a competing-risk analysis using the cumulative incidence function. The cumulative incidence function for progression at day 14 was 23.7% (95% CI: 17.8–29.7%). Compared with the Kaplan-Meier estimates, the cumulative incidence function was lower at day 14, suggesting that the Kaplan-Meier method may overestimate the absolute risk at later time points due to the high competing event rate.

Analysis of laboratory data characteristics in severe and non-severe groups of elderly hospitalized patients with influenza A virus infection

After excluding variables with > 5% missing data, 27 laboratory biomarkers measured within 24 h of admission were included in the final analysis (Table 2 and Supplementary Table S1). Sixteen of these biomarkers showed statistically significant differences (P < 0.05) between the severe and non-severe groups (Table 2), including WBC, Hb, LYM%, NEUT%, CRP, PCT, IL-6, BUN, ALB, GLU, CK-MB, MYO, TnT, PT, APTT, and D-D. The severe group exhibited significantly higher levels of 13 biomarkers compared to the non-severe group. These included inflammation-related markers (WBC, NEUT%, CRP, PCT, IL-6), cardiac markers (CK-MB, MYO, TnT), coagulation markers (PT, APTT, D-D), GLU, and BUN. Conversely, Hb, LYM%, and ALB levels were significantly lower in the severe group (Table 2).

Table 2.

The laboratory data characteristics of patients in the severe group and the non-severe group

Biomarkers, median (IQR)	Total (n = 198)	Severe (n = 50)	Non-severe (n = 148)	P-value
WBC, 109/L	6.70 (4.52–9.99)	10.80 (6.91–13.57)	5.98 (4.33–7.87)	< 0.001
Hb, g/L	122.00 (105.25–133.00)	115.00 (95.00-128.75)	123.00 (108.75–134.00)	0.005
LYM%	15.15 (8.22–23.10)	7.95 (4.05–15.88)	17.45 (10.55–24.83)	< 0.001
NEUT%	74.90 (64.70-83.88)	86.20 (77.00-91.45)	72.40 (63.98–80.40)	< 0.001
CRP, mg/L	21.08 (7.81–68.46)	90.25 (29.06-158.49)	14.62 (5.68–39.50)	< 0.001
PCT, ng/mL	0.12 (0.05–0.37)	0.66 (0.24–1.43)	0.08 (0.04–0.18)	< 0.001
IL-6, pg/mL	31.70 (13.61–78.56)	119.95 (55.20-281.82)	23.63 (12.21–50.05)	< 0.001
BUN, mmol/L	6.70 (4.82–9.51)	7.70 (5.80–17.50)	6.26 (4.56–8.89)	< 0.001
ALB, g/L	37.16 (33.61–40.82)	34.17 (30.60-38.18)	38.09 (35.00-41.36)	< 0.001
GLU, mmol/L	6.90 (5.80–8.70)	7.35 (6.23–10.20)	6.80 (5.65–8.33)	0.037
CK-MB, ng/L	1.52 (1.02–2.88)	2.88 (1.18–5.04)	1.44 (1.00-2.56)	0.002
MYO, ng/L	52.64 (33.69-102.15)	85.37 (48.48-227.08)	43.78 (31.56–77.47)	< 0.001
TnT, ng/mL	0.016 (0.011–0.037)	0.037 (0.014–0.062)	0.014 (0.011–0.025)	< 0.001
PT, s	12.20 (11.50-13.07)	12.60 (11.85–13.90)	12.10 (11.40–12.90)	0.003
APTT, s	29.90 (27.80–32.60)	30.95 (28.40-33.85)	29.90 (27.40-32.45)	0.050
D-D, mg/L	0.66 (0.46–1.36)	1.71 (0.91–4.30)	0.60 (0.40–1.03)	< 0.001

WBC, white blood cell count; Hb, hemoglobin; LYM%, lymphocytes percentage; NEUT%, neutrophil percentage; CRP, C-reactive protein; PCT, procalcitonin; IL-6, interleukin-6; BUN, blood urea nitrogen; ALB, albumin; GLU, glucose; CK-MB, creatine kinase-myocardial band; MYO, myoglobin; TnT, troponin T; PT, prothrombin time; APTT, activated partial thromboplastin time; D-D, D-dimer

Screening of predictors for severe influenza A virus infection in hospitalized elderly patients

We finally considered 29 variables as candidate predictors, including sex, age, and the 27 retained laboratory biomarkers (Table 2 and Supplementary Table S1). The dataset was split into a training set (n = 140) and a validation set (n = 58) using a 7:3 ratio. The training set was used for predictor screening and model development, and the validation set served for internal validation. To ensure the training and validation sets were balanced, we compared the distribution of each variable between the two sets. No statistically significant differences were observed (P > 0.05 for all variables; Supplementary Table S2).

Elastic net regression was then applied to the 29 candidate predictors to select features for the model. Using the one standard error (1SE) rule to promote parsimony, the algorithm identified four variables that were retained for further analysis: WBC, PCT, IL-6, and MYO (Fig. 2). These four variables were included in the Cox regression analysis. The proportional hazards assumption was satisfied for each variable (WBC: P = 0.42; PCT: P = 0.97; IL-6: P = 0.47; MYO: P = 0.39; global test: P = 0.53).

Fig. 2.

Screening predictors through elastic net regression. (A) Cross-validation error plot; (B) Coefficient path plot of each variable. As λ increases, the coefficient of each variable is gradually reduced to zero, and four variables with non-zero coefficients under the 1SE rule (red dashed line) are selected as predictors

Development of a prediction nomogram for severe influenza A virus infection in hospitalized elderly patients

We developed a prediction nomogram (Fig. 3) for severe influenza A virus infection in hospitalized elderly patients based on the Cox regression analysis (Fig. 4 and Supplementary Table S3). The nomogram uses the four predictors identified in the regression to calculate the probability of disease progression to severe illness within 14 days of hospitalization. Each variable is assigned points on the top axis; these points are summed to obtain a total score, which corresponds to the predicted probability of severe disease at 14 days.

Fig. 3.

A nomogram for predicting severe influenza based on Cox regression. The values on the scale of each variable correspond to the points in the first row. Summary results of these points are shown in the form of a total score, and the total score corresponds to the probability of severe influenza for the patient within 14 days

Fig. 4.

Forest plot of hazard ratios in the Cox model. Horizontal lines indicate 95% confidence intervals, and asterisks represent significance (*P < 0.05, **P < 0.01, ***P < 0.001)

To ensure full transparency and reproducibility, the model can also be expressed algebraically. The linear predictor (LP) for an individual patient is calculated as:

The baseline survival probabilities at day 14, estimated from the training set when LP = 0, was S₀(14) = 0.792. The absolute risk of progression to severe influenza at day 14 is then:

This formula yields exactly the same probabilities as those read from the nomogram and allows any researcher to compute individual risks without the graphical tool.

Preliminary evaluation of the prognostic model for severe influenza A virus infection in hospitalized elderly patients

We preliminarily evaluated the model’s performance. Using ROC curves, we assessed the model’s predictive performance, yielding C-index values of 0.851 in the training set and 0.819 in the validation set. The AUC values at day 14 exceeded 0.8 in both the training and validation sets (Fig. 5), suggesting acceptable discriminatory ability, although uncertainty remains due to the modest sample size and limited number of events. Bootstrap-based internal validation (1,000 resamples) yielded estimates broadly consistent with the split-sample results, although uncertainty remains given the modest sample size, limited number of events, lack of external validation, and the fact that the full modeling pipeline was not repeated in each bootstrap sample (Supplementary Figure S2). Calibration plots indicated that the model’s predicted probabilities at day 14 showed agreement with the observed incidence rates (Fig. 6). Additionally, given the modest sample size and limited number of events, all results should be interpreted with caution and considered preliminary and exploratory.

Fig. 5.

Analysis of the discriminative ability of the model. (A, B) The ROC curves and C-index values of the model in training and validation cohorts respectively. Blue: the 14-day ROC curve. (C, D) The time-dependent AUC curves of the model in training and validation cohorts displaying the trend of the model’s discriminative ability over time. Blue shadow and red shadow: 95% confidence intervals

Fig. 6.

Calibration curves illustrating the relationship between predicted probabilities of the model and actual occurrence rates. (A) Calibration plots at day 14 for the training cohort. (B) Calibration plots at day 14 for the validating cohort

We additionally estimated the 7-day risk of progression to severe influenza. The nomogram, ROC curves, and poorly calibrated calibration curves are presented in Supplementary Figure S3, Supplementary Figure S4, and Supplementary Figure S5, respectively. These 7-day estimates showed suboptimal calibration and are provided only for transparency and exploratory purposes. For clarity, a summary of the key performance metrics for the training and validation sets is presented in Supplementary Table S4. Finally, exploratory analysis of sensitivity and specificity at different thresholds is provided in Supplementary Table S5. These threshold-specific estimates are purely descriptive and should not be used to guide clinical decision-making.

Discussion

Influenza A virus infections often cause nonspecific, cold-like symptoms that are easily mistaken for a common cold. This may cause patients or clinicians to underestimate the infection, leading to missed opportunities for timely antiviral treatment and an increased risk of severe illness or death. This is especially true for elderly patients with underlying health conditions, who are at a significantly higher risk of severe complications and mortality [26]. Despite numerous studies examining risk factors for severe influenza, prediction tools remain notably lacking for assessing disease progression in elderly patients infected with influenza A virus [27].

This study analyzed the clinical and laboratory characteristics of elderly hospitalized patients with influenza A virus. We then developed a laboratory‑based prognostic model based on laboratory biomarkers to predict progression to severe influenza, using a strategy combining elastic net regression with Cox regression. The resulting nomogram model incorporates four clinical biomarkers (WBC, PCT, IL-6, and MYO) identified as significant predictors. All of these biomarkers are routinely measured within 24 h of admission as part of standard care. They are generated by automated analyzers and reported in the electronic medical record, requiring no specialized data handling or additional training for collection. This ready availability makes them particularly suitable for nomogram development. Nomograms are widely used visualization tools for risk prediction in medicine, including oncology and disease risk assessment [28–31]. In this study, the nomogram serves as an exploratory tool for risk estimation, not a validated instrument for clinical decision-making.

Model performance evaluation showed moderate discrimination in both the training (C-index = 0.851; AUC = 0.855 at day 14) and validation sets (C-index = 0.819; AUC = 0.809 at day 14). Bootstrap-based internal validation (1,000 resamples) yielded comparable estimates. The 95% CIs for the 14‑day AUC (0.764–0.947), and C‑index (0.787–0.887) were broadly consistent with the split-sample results. Calibration analysis indicated potentially reasonable calibration at day 14. In summary, these findings suggest that the model may provide some exploratory information for mid-term risk assessment, but prospective validation is needed.

We selected day 14 as prediction time points based on the observed event-time distribution in our cohort, in which 30.1% of progressions occurred by day 14. These data indicate that a substantial proportion of events occurred within the first two weeks of hospitalization. We acknowledge that hospital discharge is a competing event because discharged patients are no longer at risk of in-hospital progression. The Kaplan-Meier method may overestimate the absolute risk of progression when the competing event rate is high. Indeed, our competing-risk analysis gave a lower 14-day risk (23.7%) than the Kaplan-Meier estimate (30.1%), consistent with this bias. However, the model’s discriminative ability (AUC and C-index), which depends on the ranking of risk scores rather than absolute risk values, is generally not affected by this issue [32, 33]. Importantly, the model estimates only in-hospital progression risk rather than absolute post-admission progression risk. Future studies with larger samples should consider competing-risk frameworks to obtain unbiased absolute risk estimates.

The innate immune system serves as the body’s first line of defense against influenza viruses, typically eliminating the pathogen through coordinated immune responses. However, a dysregulated host immune response can trigger a “cytokine storm,” a key factor in influenza severity and mortality. This storm can cause severe lung injury, multiple organ failure, shock, and secondary infections [34–36].

WBCs are the body’s primary immune cells, critical for resisting pathogens and promoting immune responses. In uncomplicated influenza virus infections, WBC levels usually remain normal or only slightly decrease [37, 38]. In this study, WBC levels were significantly elevated in severe influenza patients and emerged as an important risk predictor for disease progression. This finding is consistent with previous studies suggesting that elevated WBC levels may be associated with damage to the respiratory barrier and alterations in the local immune microenvironment following influenza virus infection, thereby increasing the risk of secondary infections [39–43]. Therefore, in this model, elevated WBC levels, particularly neutrophils, suggest a possible secondary bacterial co-infection rather than a direct result of the influenza virus infection.

Prior research indicates that severe cases of influenza A virus (H1N1) infection involve a rapid release of IL-6. Significantly elevated serum IL-6 levels are recognized as an important biomarker for the severity of acute lung injury [44, 45]. Of the four predictors for severe influenza A virus infection identified in our study, we incorporated IL-6—a key inflammatory factor—into the prediction model, consistent with its role as a predictive biomarker.

PCT is a protein synthesized by thyroid C-cells, with greater specificity for bacterial infections and normally minimal serum levels [46–49]. Studies have shown that elevated PCT levels often indicate secondary bacterial infections complicating viral illnesses, potentially worsening clinical symptoms. This is frequently observed in children and older adults (≥ 65 years) with viral lower respiratory tract infections [50–52]. Clinical evidence shows that monitoring PCT levels may provide information for predicting complications and disease progression in patients with viral infections [47]. Therefore, clinically significant elevations in WBC counts or PCT levels may suggest bacterial co-infection and can serve as supportive information for risk assessment.

MYO, an important biomarker in the cardiac enzyme profile, is crucial for early detection and severity monitoring of myocardial injury [53, 54]. In addition to common respiratory sequelae such as pneumonia and respiratory failure, influenza virus infection is associated with myocardial injury in approximately 10% of cases, typically occurring 4–7 days after symptom onset. The clinical presentation of influenza-associated myocardial injury is highly variable, spanning from subclinical involvement to fulminant myocarditis [55–57]. In our study, patients with severe influenza exhibited significantly elevated levels of both MYO and IL-6. These findings align with previous reports of myocardial involvement and inflammatory dysregulation in severe influenza and are consistent with the predictive value of these biomarkers in the present model [58, 59]. Furthermore, myocardial injury shows correlations with WBC and PCT levels, where elevated PCT may reflect the severity of myocardial damage [60, 61]. Thus, combined detection of WBC, PCT, IL-6, and MYO may offer preliminary information for risk assessment and hypothesis generation.

Our study has several limitations

First, as a single-center retrospective analysis focused on elderly hospitalized patients with influenza A virus infection, the generalizability of our findings is limited. The model’s performance in younger populations has not been assessed. Moreover, the modest sample size further limits the robustness of our findings. All results should therefore be interpreted with caution and considered preliminary and exploratory. The model has not been externally validated, and its value remains uncertain. Future research should include larger, independent, multi-center cohorts to validate our conclusions and evaluate the model’s applicability across different age groups and healthcare settings.

Second, the absence of certain laboratory data, individual variations in viral shedding, and variations in treatment decisions (e.g., timing of antiviral therapy, use of corticosteroids or immunomodulators) may introduce confounding biases in this retrospective analysis.

Third, our approach of excluding variables with > 5% missing data may have excluded clinically relevant predictors. Future studies with larger samples should consider retaining such variables for analysis.

Fourth, missing data were handled by averaging imputed values across five imputed datasets. This approach was used because the missing rate was low (< 5%), but we acknowledge this limitation. It does not fully reflect imputation uncertainty. Future studies may consider a more rigorous approach to better capture imputation uncertainty.

Fifth, despite the bootstrap-based internal validation yielding estimates broadly consistent with the split-sample results, this bootstrap procedure did not repeat the full modeling pipeline (including missing data handling, elastic net variable selection, and Cox regression fitting). Therefore, it did not account for the uncertainty introduced by variable selection and model fitting. The results should therefore be interpreted with caution and considered preliminary and exploratory.

Sixth, this study developed a laboratory-based prognostic model for severe influenza using routine laboratory biomarkers. Although age and sex were evaluated as candidate predictors, only laboratory biomarkers were retained in the final model after elastic net selection. Moreover, several comorbidities (including heart disease, liver disease, renal disease, and brain disease) differed significantly between severe and non-severe patients, but these clinical variables were not evaluated as candidate predictors in the elastic net selection. As a result, the model’s performance relative to basic clinical predictors or existing clinical judgement has not been evaluated, and whether the model adds incremental value beyond simple clinical information remains to be determined by systematic comparisons of clinical, laboratory, and combined models in future large-scale studies. Therefore, the model should be viewed as exploratory in nature, not as a clinically validated tool. This limitation should be considered when interpreting the results.

Conclusion

In summary, we have developed and preliminarily evaluated a laboratory-based prognostic model (WBC, IL-6, PCT, MYO) to predict in-hospital progression to severe disease in elderly patients with influenza A virus infection. The model showed moderate discriminative ability in this single-center cohort. However, given the study’s limitations, all findings should be interpreted as exploratory. Further studies are necessary to validate these findings and to evaluate the model’s reproducibility and potential value.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

NPIs

Non-pharmaceutical interventions

WBC

White blood cell count

Hb

Hemoglobin

LYM%

Lymphocyte percentage

NEUT%

Neutrophil percentage

CRP

C-reactive protein

SAA

Serum amyloid A

PCT

Procalcitonin

IL-6

Interleukin-6

sCr

Serum creatinine

BUN

Blood urea nitrogen

TP

Total protein

ALB

Albumin

TBIL

Total bilirubin

ALT

Alanine aminotransferase

AST

Aspartate aminotransferase

ALP

Alkaline phosphatase

GGT

Gamma-glutamyl transferase

GLU

Glucose

K⁺

Potassium ion

Na⁺

Sodium ion

Cl⁻

Chloride ion

CK-MB

Creatine kinase-myocardial band

MYO

Myoglobin

TnT

Troponin T

NT-proBNP

N-terminal pro-B-type natriuretic peptide

BNP

B-type natriuretic peptide

PT

Prothrombin time

APTT

Activated partial thromboplastin time

TT

Thrombin time

FIB

Fibrinogen

D-D

D-dimer

FDP

Fibrin(ogen) degradation products

SaO₂

Oxygen saturation

ROC

Receiver operating characteristic curve

AUC

Area under the receiver operating characteristic curve

C-index

Concordance index

CI

Confidence interval

LP

Linear predictor

Author contributions

LX collected data and wrote the first draft. FBY performed the data analysis. ZZF provided technical support for the graphics. ZNN and ZH assisted with data collection. MQH designed the experiments, reviewed the first draft, and approved the final manuscript. All authors read and approved the final manuscript.

Funding

This study was supported by the Scientific Research Start-up Special Fund of Shanghai Fourth People’s Hospital, affiliated with Tongji University (sykyqd02201), the Discipline Boosting Program of Shanghai Fourth People’s Hospital, affiliated with Tongji University (SY-XKZT-2025-2001), and the Scientific Research Project of the Shanghai Municipal Health Commission (202240063). None of the funders were involved in the research design, data collection and processing, publication and writing of the manuscript.

Data availability

Anonymized data and analytical code are available from the corresponding author upon reasonable request, subject to ethics approval where required.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of Shanghai Fourth People’s Hospital, affiliated with Tongji University (No. 2025079-001) and retrospectively registered with the Chinese Clinical Trial Registry (ChiCTR2500112511) on November 14, 2025. The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. The Ethics Committee of Shanghai Fourth People’s Hospital, affiliated with Tongji University, waived informed consent for all patients in this retrospective study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.