Study population

An initial dataset of 519,705 records from January to September 2024 was extracted. After excluding records involving co-detection to minimize attribution bias that may confound pathogen-specific analyses, along with those containing incomplete information, duplicate same-day tests, or abnormal data, a total of 421,776 test results were retained for final analysis (Fig 1). These records were derived from individuals residing in Beijing (169,289, 40.1%), Chengdu (5,964, 1.4%), Guangzhou (41,524, 9.8%), Hangzhou (8,422, 2.0%), Nanjing (2,843, 0.7%), Shanghai (160,132, 38.0%), and Shenzhen (33,602, 8.0%). The median age of tested individuals was 30 years (interquartile range: 8–38 years), with females accounting for 57.0%. Age group distribution was as follows: infants (0 years), 1.7%; children (1–5 years), 15.1%; school-age adolescents (6–17 years), 22.3%; adults (18–59 years), 54.9%; and older adults (≥60 years), 6.0%. The majority of tested individuals were concentrated in the early months of the year, with January alone accounting for 29.9% of the total. The number of tested individuals then declined steadily over the months, reaching the lowest level in September, which accounted for only 5.4% of the total.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Data processing flowchart and baseline characteristics of the analytic data.

The flow diagram summarizes inclusion and exclusion steps applied to the initial 519,705 records from January to September 2024, yielding a final analytic set of 421,776 single-pathogen infection samples with complete symptom data. The bottom panels show the characteristics of the final analytic cohort by age group, sex, and month of specimen collection; values are counts with percentages, reflecting the distribution among the 421,776 included samples.

https://doi.org/10.1371/journal.pdig.0001656.g001

Pathogen positivity patterns

In total, 37.9% (159,881/421,776) of participants were identified as single-pathogen positive, with SARS-CoV-2 (11.2%), Human adenovirus (HAdV) (5.5%), and Human rhinovirus (HRV) (5.0%) emerging as the three most frequently detected pathogens among the nine tested during the study period (S1 Table in Supporting Information). Clear seasonal dynamics were observed (S1 Fig in Supporting Information): Influenza B was initially the predominant virus in January; however, with the rapid rise in SARS-CoV-2 infections, the latter soon became the leading pathogen, reaching its first peak in March. After this initial peak, SARS-CoV-2 declined in March but resurged in late June, reaching a second peak in August. These trends are closely aligned with the temporal patterns of SARS-CoV-2 and influenza virus positivity rates among influenza-like illness (ILI) cases reported by national sentinel hospitals, as published in the Chinese Center for Disease Control and Prevention (China CDC) surveillance data [10]. Biweekly positivity rates of 9 tested respiratory pathogens stratified by sex and age group are available in S2 Fig and S3 Fig in Supporting Information.

Pathogen positivity rates varied notably by sex and age group (S1 Table in Supporting Information). Females exhibited significantly higher positivity rates for SARS-CoV-2 and respiratory syncytial virus (RSV), but lower rates for the other seven pathogens compared to males. Age-related differences were also observed. For example, the positivity of SARS-CoV-2 was markedly higher among older adults (≥60 years), while the positivity of HAdV peaked among young children (1–5 years) and school-aged individuals (6–17 years).

Symptom signature of pathogen stratified by age and sex

We compared 13 clinical symptoms among patients infected with nine respiratory pathogens and further stratified by age and sex (Table 1 and S4 Fig in Supporting Information). Across pathogens, fever and cough were the most frequently reported symptoms, and overall symptom signatures showed broad overlap. However, differences in symptom distribution and severity were observed. For example, RSV, Human coronavirus (HCoV), Human parainfluenza virus (HPIV), and HRV were predominantly associated with respiratory symptoms, whereas influenza A, influenza B, and SARS-CoV-2 also showed more pronounced systemic symptoms, such as fever, headache, and myalgia. Moreover, HRV presented the lowest fever prevalence and rarely exhibited severe fever phenotypes (e.g., high-grade fever ≥39 °C or fever ≥3 days), whereas high-grade fever occurred more often in HAdV and influenza A.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Age-specific symptom patterns in patients with single-pathogen infections.

https://doi.org/10.1371/journal.pdig.0001656.t001

Age-stratified analysis revealed that, across the nine pathogens, most symptom distributions varied across age groups. Compared with adults aged 18–59 years, children and school-age adolescents (0–17 years) exhibited higher frequencies of fever-usually with greater severity and relatively lower frequencies of cough for most pathogens. Adults ≥60 years showed symptom distributions broadly similar to those of the 18–59 year group, but tended to report cough more often across many pathogens. Morever, symptoms such as sore throat, rhinorrhea, nasal congestion, and headache were generally less frequent in the 0–17 and ≥60-year groups than in the 18–59-year group. Sex-stratified analyses (S4 Fig in Supporting Information) revealed significant differences for certain symptoms, with the symptoms involved and the direction of effect varying by pathogen. Notably, high-grade fever (≥39°C) occurred more often in males than in females for nearly all pathogen groups, with the exception of M. pneumoniae.

Overall, symptom signatures largely overlapped across pathogens; observed differences were more categorical than pathogen-specific. Moreover, symptom expression was further influenced by age and sex.

Symptom-based hierarchical clustering of respiratory pathogens and development of predictive models

Based on the similar features observed across some age groups, we further combined them in the main analysis into three categories: 0–5 years, 6–17 years, and ≥18 years, while the more detailed age stratifications were incorporated into the sensitivity analysis. In adults≥18 years, hierarchical clustering grouped the nine pathogens into three stable classes with similar symptom signature: Class 0-influenza A, influenza B, and SARS-CoV-2; Class 1-HAdV and M. pneumoniae; and Class 2-HCoV, HPIV, HRV, and RSV (Table 2). These clustering patterns were highly consistent across adult subgroups (18–45, 45–59, and ≥60 years). Clustering pattern in younger age groups showed some divergence from the adults group: in the 6–17 year group two pathogens were reassigned, with M. pneumoniae clustering with the influenza viruses and SARS-CoV-2 clustering with HAdV (Table 2). However, 0–5 years showed markedly different results (details provided in S5 Fig in Supporting Information), likely due to limited symptom self-reporting in this age group.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Pathogen clustering based on symptom distribution and prediction performance of the MLP model.

https://doi.org/10.1371/journal.pdig.0001656.t002

We then trained a MLP model to predict the broad pathogen classes from symptoms (Table 2). Although the clustering was entirely data-driven, certain classes showed relatively consistent symptom patterns (Fig 2), which were highly aligned with the features identified earlier section. For example, in both the 6–17 and ≥18 years age groups, HCoV, HPIV, HRV, and RSV were clustered together, and these infections tended to produce typical upper respiratory symptoms such as rhinorrhea, cough, and sore throat, with a relatively lower prevalence of systemic symptoms compared with the other classes. In contrast, infections with influenza A, influenza B, SARS-CoV-2, and M. pneumoniae-although the clustering of SARS-CoV-2 and M. pneumoniae varied slightly across age groups—were consistently characterized by more systemic manifestations and higher rates of fever.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Symptom-based pathogen class definition and MLP classification performance stratified by three age groups.

The left panels show hierarchical clustering of nine pathogens into three classes (labeled 0, 1, and 2 at the top) based on their symptom profile distributions within each of three age groups. Each column represents a specific pathogen and each row represents a symptom, with the color gradient from blue to red indicating increasing prevalence of the symptom in infections by specific pathogen. The right panels depict the performance of a MLP model trained to classify individual infections into these three predefined classes: one-vs-rest ROC curves are shown for each age group, with associated AUC values annotated, reflecting the model’s ability to differentiate symptom-defined infections across pathogen classes.

https://doi.org/10.1371/journal.pdig.0001656.g002

The MLP model demonstrated overall good ability to distinguish broad pathogen classes, broad pathogen classes, with area under the curve (AUC) values of 0.85-0.91 in adults≥18 years, 0.86-0.92 in infants and children aged 0–5 years, and slightly lower values of 0.80-0.83 in those aged 6–17 years (Table 2). Calibration plots suggested reasonable agreement between predicted and observed probabilities (S6 Fig in Supporting Information). Notably, the model exhibited a conservative classification tendency across most class categories, with specificity all above 75%, but relatively lower sensitivity in several classes. For example, in the ≥ 18 years group, Class 1 (HAdV and M. pneumoniae) had a sensitivity as low as 57.7%. In the 0–5 years group, Class 1 (M. pneumoniae) had a sensitivity of only 53.1%. The high AUCs and specificities but low sensitivities indicate a tendency to under-detection. Incorporating epidemiological information may help improve case identification.

Refining predictions of each pathogen by integrating epidemiological data

Fig 3 illustrates the comparison between the post-hoc reweighted MLP predictions–adjusted using pathogen epidemiological data–with the observed pathogen distributions across age groups. When only the top-ranked predicted pathogen was used as the final classification, the overall accuracy was only 0.45 [95% CI, 0.43-0.46] (κ = 0.32 [95%CI, 0.31-0.32]) in 6–17 years group and 0.49 [95% CI, 0.48-0.51] (κ = 0.39 [95%CI, 0.37-0.39]) in individuals ≥18 years. Regarding prediction of individual pathogens, the model achieved moderate accuracy for some pathogens (S2 Table in Supporting Information). For example, in the 6–17 year old age group, the model achieved a sensitivity of 0.64 (95% CI: 0.61-0.67) and a specificity of 0.71 (95% CI: 0.70-0.73) for detecting HAdV. Among adults aged≥18 years, the model demonstrated high sensitivity for identifying SARS-CoV-2, reaching 0.88 (95% CI: 0.87–0.90). However, the model demonstrated little to no capacity to distinguish certain pathogens—including HCoV, HPIV, influenza A, and RSV—with sensitivities approaching zero. Expanding the predicted target to include the top two most probable pathogens improved overall performance. The classification accuracy increased to 0.67 (95%CI, 0.66-0.69) in aged 6–17 and 0.71(95%CI, 0.70-0.72) in those ≥18 years. Further extending the candidate range to the top three predictions resulted in an accuracy of 0.83 (95%CI, 0.82,0.84) for 6–17-year age group and 0.83 (95%CI, 0.82,0.83) for ≥18-year-olds. Overall, integration of epidemiological data enhanced predictions for prevalent pathogens but not for less common ones; however, broadening the prediction to include the top two or three likely pathogens markedly improved overall accuracy.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Confusion matrix of MLP top-1 pathogen predictions after integration of epidemiological data.

This figure shows the confusion matrix of top-1 pathogen predictions in the 6–17 and 18+ years-old age groups, after post-hoc reweighting of MLP model outputs with epidemiological data. Each cell shows both the number of samples and the percentage of all cases with that true pathogen that were predicted as the corresponding pathogen. Cells are shaded from white to blue according to the prediction distribution percentage, with darker indicating higher proportions. Red outlines highlight the diagonal cells representing correct predictions.

https://doi.org/10.1371/journal.pdig.0001656.g003

Ethics statement

The study protocol was reviewed and approved by the Ethics Committee of China-Japan Friendship Hospital (Approval No. 2024-KY-436). The data used in this study were obtained through a research collaboration with Meituan Healthcare (Beijing, China), which provided anonymized respiratory specimen and pathogen testing data under a national research project for research only. Meituan had no role in the study design, data interpretation, or manuscript preparation. All participants provided electronic informed consent for the use of anonymized testing data for scientific research purposes. For the present study, only fully anonymized secondary data were analyzed, and no identifiable personal information was accessed by the researchers.

Study design and data source

A cross-sectional study was conducted based on data from Meituan Healthcare (Beijing, China), a nationwide pharmaceutical e-commerce platform. The study population comprised individuals who self-purchased multiplex respiratory pathogen (MRP) testing kits due to self-perceived upper respiratory tract infection symptoms and opted for self-testing via the platform. Before ordering the testing kits, participants were required to provide informed consent for the use of anonymized testing data in scientific research and complete a brief questionnaire collecting information on sex, age, and symptoms. Reported symptoms included fever, fever≥ 39 °C, fever lasting≥ three days, cough, sputum production, sore throat, rhinorrhea, nasal congestion, dysphonia, headache, myalgia, diarrhoea, and vomiting. For young children unable to complete the ordering process independently, parents could purchase the testing service on their behalf, provide the necessary personal information, and assist with sample collection. The samples were transported to laboratories accredited by the China National Accreditation Service for Conformity Assessment (CNAS) by trained delivery personnel. Polymerase chain reaction (PCR) testing was conducted to simultaneously detect nine pathogens: influenza A, influenza B, RSV, HRV, HPIV, HAdV, HCoV, M. pneumoniae, and SARS-CoV-2.

Data management and statistical analysis

The study included participants who conducted self-testing, provided informed consent, and received test results between January 1 and September 30, 2024. Their test results, along with self-reported sex, age, and symptoms from the questionnaire, were extracted from the Meituan Healthcare platform. Due to the accessibility of test kit purchases, participants were primarily from seven major cities in China: Beijing, Chengdu, Guangzhou, Hangzhou, Nanjing, Shanghai, and Shenzhen. Individuals with missing symptom data or test results were excluded. Additionally, to minimize the influence of co-infections on symptom pattern analysis, individuals with multi-pathogen infections were excluded.

Five age groups were defined for analysis: infants (0 years), children (1–5 years), school-age individuals (6–17 years), adults (18–59 years), and older adults (≥60 years). We examined changes in the detection rates of upper respiratory tract pathogens across different age and sex groups over the study period, defined as the proportion of individuals with single-pathogen infections among all tested individuals. We restricted the symptom analysis to single-pathogen infections and compared symptom profiles across pathogens. Individuals with co-detection were excluded to avoid attribution bias, which can confound pathogen-specific comparisons and degrade model performance. Categorical variables were summarized using counts and percentages. The chi-square test was employed to assess differences in qualitative variables, and differences between proportions were evaluated using the two-sample proportion test. P values were adjusted for multiple comparisons using the Bonferroni method, and a two-sided P value of <0.05 was considered statistically significant.

Symptom-based clustering and MLP modeling

To simplify the classification task and investigate the feasibility of symptom-based pathogen prediction, we first applied hierarchical clustering to group the nine pathogens into three broader classes based on their age-stratified symptom distribution patterns. This class-merging strategy led to improved classification performance in preliminary experiments, outperforming direct nine-class classification models in accuracy and stability. For the three-class classification task, we developed an MLP model comprising three fully connected layers. Each layer was followed by a ReLU activation function to introduce non-linearity, and a dropout layer with a dropout probability of 0.5 to prevent overfitting. The model was optimized using the Adam algorithm with an initial learning rate of 1e-4, β₁ = 0.9, β₂ = 0.999, and a weight decay of 0.001.

To address class imbalance in the training data, we employed a k-means–based downsampling strategy, where samples from the majority classes were clustered and downsampled to match the size of the minority class, ensuring a balanced dataset across all categories. The final dataset of single-pathogen infection cases was randomly split into training (60%), validation (20%), and test (20%) sets. Model training was conducted for a maximum of 100 epochs, and the model with the lowest validation loss was selected as the final model. Performance on the test set was evaluated using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), with 95% confidence intervals estimated using bootstrap resampling. Model calibration was assessed by calibration plots comparing predicted probabilities with observed event proportions for each class across age groups.

Integrating epidemiological data to enhance MLP predictions

To assess whether incorporating epidemiological data can improve pathogen identification, we performed a post-hoc refinement of MLP predictions. Based on the predicted probabilities of the three broader pathogen classes generated by the MLP, we first assigned equal probabilities to each pathogen within a given class. These initial probabilities were then adjusted using the two-week detection rates of each pathogen at the population level. The weighted probabilities were normalized to sum to one for each case, and the pathogen with the highest adjusted probability was taken as the predicted result. The pathogen with the highest final probability was taken as the predicted result. Prediction performance was assessed using multiple complementary metrics. Overall agreement between predicted and true pathogen classes was quantified by classification accuracy and Cohen’s Kappa coefficient, while per-pathogen diagnostic performance was characterized by sensitivity, specificity, precision, and F1-score, each with 95% confidence intervals. Statistical analysis was performed with R software, version 4.3.2.