Trends and factors influencing over-the-counter cold and cough medication sales in Guangzhou, China: 2011–2017

Abstract

Background

People are inclined to resort to Over-the-counter cold and cough medications (OTCCM) for self-treatment when experiencing upper respiratory tract infections, with consumption habits varying significantly among residents across different regions. Fluctuations in pharmacy OTCCM sales stem from numerous and intricate factors. However, most current studies predominantly examine the impact of singular factors on OTCCM volume. A comprehensive quantitative understanding of the evolving trends in OTCCM sales and the collective influence of multiple factors remains incomplete, particularly in subtropical mega-urban areas characterized by unique seasonal patterns of infectious diseases.

Methods

We monitored OTCCM sales at 36 stores of a major pharmaceutical chain in Guangzhou, China, from 2011 to 2017. We explored the association between OTCCM% and meteorological, infectious disease, and population movement factors using wavelet coherence and cross-correlation analysis. Extreme gradient boosting and SHapley Additive exPlanations were used to characterize each factor’s contributions.

Results

The OTCCM% in Guangzhou exhibited significant seasonal and annual cycles, with peaks occurring in the winter-spring and summer seasons. Meteorological factors, infectious disease factors, and OTCCM% were correlated within an annual cycle. OTCCM% had the highest correlation with the proportion of influenza-like illness at the same time or one week earlier during the winter-spring peaks. We revealed the contribution of infectious disease factors (38.15%), population movement factors (38.11%), and meteorological factors (23.74%) on OTCCM%, explicitly identifying the top three contributing factors as mobile population, influenza B, and mean temperature. The effect curve of mean temperature was reversed J-shaped, while the curves for relative humidity, sunshine duration, and outpatient visits were V-shaped. As influenza and other upper respiratory infections increased, OTCCM% showed an upward trend. Increased mobile population and in-migration index also contributed to the rise in OTCCM%.

Conclusion

This study comprehensively assesses the impact of meteorological, infectious disease, and population movement factors on OTCCM sales in a subtropical mega-city. These insights can inform local consumer behaviors and public health strategies, aid in business market forecasting, and support the allocation of medication resources and the prediction of disease outbreaks.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-025-11247-0.

Background

Over-the-counter cold and cough medications (OTCCM) have become widely accepted in public health, representing a crucial category in the global over-the-counter (OTC) market. When individuals experience upper respiratory tract infections, there is a common tendency to choose non-prescription drugs over seeking immediate medical attention [1, 2]. The sale of pharmacy-based OTCCM is essential for providing convenient public health services [3, 4]. Health policies and residents’ OTCCM consumption habits vary significantly across regions. The rising demand for medications during public health emergencies in recent years suggests that a quantitative understanding of medication sales trends and the factors influencing them remains necessary, particularly for megacities with unique seasonal patterns of infectious diseases (e.g., Guangzhou, China) [5–7].

Medication sales data have shown significant associations with the prevalence of infectious diseases and hold potential for the early detection of outbreaks [7]. Studies conducted in France and New York City have demonstrated correlations between the sales of over-the-counter (OTC) medications and the incidence of influenza-like illness (ILI) and gastrointestinal disease [8, 9]. In Tianjin, China, a syndromic surveillance study reported a correlation coefficient of 0.49 between OTC sales and laboratory-confirmed influenza positivity rates [3]. Existing research has also examined the impact of meteorological changes on the sales of medications. A study conducted in five Spanish cities found a non-linear inverted J- or V-shaped positive correlation between the number of medically prescribed drugs for respiratory diseases and the average daily temperature [10, 11]. However, no studies have simultaneously examined the combined effects of meteorological factors and disease prevalence on the sales of medication.

In real-world settings, fluctuations in OTCCM sales are likely driven by a complex interplay of multiple factors [9]. In addition to temperature, previous studies have identified the roles of relative humidity and sunshine duration in influencing the prevalence of infectious diseases, which may impact medication purchasing behavior [12].

Existing studies have been conducted in areas with population sizes ranging from 100,000 to several million; however, few have focused on megacities in subtropical regions with populations exceeding 10 million [10, 11, 13]. Mega-cities are often characterized by rapid economic development, increased population density, greater mobility, and more frequent interpersonal contact—all of which contribute to a significantly heightened potential for viral transmission [14]. Population migration and mobility patterns can directly alter the trajectory of infectious disease transmission. Seasonal population movements, such as those occurring during holidays, can lead to shifts in regional disease burdens and subsequently impact the allocation of public health resources [15, 16]. Furthermore, factors related to infectious diseases (such as ILI activity, influenza outbreaks, and hospital outpatient volumes) may directly influence OTCCM sales. The seasonal pattern of influenza in subtropical regions differs from that in temperate zones, typically exhibiting a bimodal distribution with peaks in both winter-spring and summer. Therefore, when conducting studies in subtropical megacities, it is essential to include a broad range of influencing factors for a more comprehensive analysis.

Meanwhile, existing studies on factors influencing medication sales have typically focused on a single variable within a specific category. The combined effects of multiple influencing factors on OTCCM sales require further exploration. Located on the southern coast of China, Guangzhou City has a population of over 12 million and serves as a major international transportation hub. This study aims to characterize OTCCM sales trends at the city level systematically, assess the impact of meteorological factors, infectious disease factors, and population movement factors on OTCCM sales in Guangzhou City, and quantify the contributions of various factors and their complex effect relationships using machine learning models. Through OTCCM sales data, we could gain insights into local consumer behaviors and their public health coping mechanisms, which will aid businesses in forecasting the OTCCM market to ensure supply-demand equilibrium. Simultaneously, it can provide scientific evidence for public health departments to forecast disease outbreaks, formulate health policies, and allocate resources to ensure the security of medication quantities.

Methods

Setting

This study was conducted in Guangzhou, a first-tier city and international comprehensive transportation hub, as well as the capital of Guangdong Province. It is located at 112°57’E to 114°3’E and 22°26’N to 23°56’N and covers an area of 7434.40 km². Situated in southern China, it experiences an oceanic subtropical monsoon climate. The permanent resident population was 14.49 million at the end of 2017. The average temperature in Guangzhou ranges from 20 to 22 degrees Celsius, with a relative humidity of approximately 77%. From December through April, Guangzhou experiences a winter-spring season (49 to 16 weeks), while the hot summer season lasts from May to October (17 to 44 weeks). November brings a short fall season with moderate temperatures (45 to 48 weeks).

Data sources

Considering previous studies and Guangzhou’s characteristics, we examined the impact of three factors on the sales of over-the-counter cold medicines in Guangzhou: infectious disease factors, meteorological factors, and population movement factors.

Medication sales data

OTCCM%

Thirty-six stores of a major pharmacy chain in Guangzhou, covering six major urban areas, were monitored for sales. Over-the-counter medications primarily used to treat cold-related symptoms, such as fever, sore throat, and cough, were selected based on market sales volume and the medication habits of residents. The study collected data on medication sales from 2011 to 2017, including sales figures of the monitored medications (measured in the smallest unit of sale) and the total sales of all medications in pharmacies during the same time frame. OTCCM% is the current week’s OTCCM sales divided by OTC sales. We used OTCCM% to represent OTCCM sales.

Infectious disease factors

Infectious disease factors are epidemiologic indicators that primarily include data from the Chinese National Influenza Surveillance Information System (CNISIS). In 2014, an outbreak of dengue fever occurred in Guangzhou, and we also assessed the impact of this public health emergency.

Outpatient visit

Data from the Guangzhou Center for Disease Control and Prevention (Guangzhou CDC) indicate the weekly outpatient visits to sentinel hospitals for influenza surveillance in Guangzhou.

ILI%

Influenza-like illness surveillance data were acquired from the Guangzhou CDC. Patients with a body temperature of 38 °C or higher and experiencing symptoms of cough or sore throat were defined as having ILI. The weekly proportion of ILI cases (ILI%) was calculated by dividing the number of outpatient visits to sentinel hospitals by the total number of outpatient visits.

Influenza epidemic intensity

Following the receipt of specimens from sentinel hospitals, the Guangzhou Influenza Surveillance Network Laboratory conducted nucleic acid assays to determine the subtype or spectrum of the influenza virus within three days. The positivity rate (PR) for influenza subtypes was established by computing the ratio of positive cases of various influenza subtypes per week to the total number of laboratory tests conducted during that particular week. PR for different influenza subtypes multiplied by ILI% was used to reflect influenza epidemic intensity and incorporated into the modeling constructs. Untyped strain A data were excluded, while A/H3N2 and A/H1N1 (including A/H1N1pdm09) were retained (A/H1N1, A/H3N2). B/Victoria, B/Yamagata, and untyped B were collectively referred to as influenza B (B).

ILI exclude flu

ILI% identified as not being influenza were separated from the ILI% and counted as a new variable called ILI exclude flu. This new variable represents upper respiratory infections similar to influenza.

Dengue

Following a record-breaking outbreak of dengue fever in Guangzhou in 2014, which lasted 193 days from June 11 to December 21 and resulted in 37,376 cases, the city implemented a range of community-based interventions to control the outbreak [17, 18]. Based on the peak of the dengue outbreak and the timing of the community-based measures, we assigned a value of 1 to the dengue variable (dengue) from August to December 2014 and 0 to the rest of the time.

Meteorological factors

The meteorological data of Guangzhou was obtained from the China Meteorological Data Service Center (http://data.cma.cn/en, last accessed: 2024-04−17). The data was organized into weekly average temperature (mean temperature), weekly average relative humidity (relative humidity), weekly precipitation (precipitation), and weekly hours of sunshine (sunshine duration).

Population movement factors

As a megacity with a resident population exceeding 12 million, Guangzhou has witnessed a significant increase in its mobile population, driven by rapid economic and social development, as well as the expanding scale of population movement between urban and rural areas, and between cities and towns. To respond to the population movement in Guangzhou, we used the long-term indicator ‘mobile population’ and the short-term indicator ‘migration index.’

Mobile population

The mobile population in Guangzhou from 2011 to 2017 was calculated by subtracting the registered population from the permanent resident population, based on data from the Guangdong Provincial Statistical Yearbook. In this context, the permanent resident population refers to individuals who have lived in Guangzhou for more than six months, while the registered population includes those who hold official household registration (hukou) in the city. Therefore, the mobile population comprises individuals residing in Guangzhou without local hukou registration, reflecting the city’s ability to attract non-registered residents for employment and living purposes. The annual mobile population data were then converted to weekly estimates using linear interpolation [19, 20]. In Guangzhou, the mobile population is primarily driven by long-term economic factors and is less affected by short-term seasonal fluctuations. Linear interpolation is widely applied in macro-level demographic analyses where population trends are relatively stable. Mobile population responds to long-term population movement trends.

In and out migration indices

The researchers acquired the 2019 Guangzhou municipal GaoDe migration index from the GaoDe map (AMAP) website (https://trp.autonavi.com/index.do, last accessed: 2024-04−17), which comprises the daily in and out-migration indices (in-migration index, out-migration index). The migration index is not an actual number of population movements. However, it is proportional to the actual number, and this variable indicates short-term trends in weekly population movement. Using Eqs. 2 − 1 and considering mobile population data and statutory holidays, we extrapolated the migration index from 2011 to 2017.

 

2-1

 represents the migration index corresponding to week w of year y,  represents the annual mobile population in year y, obtained from the same source as described in the “Mobile population” section (i.e., Guangdong Provincial Statistical Yearbook). We assume that the weekly distribution pattern of population movement in 2019 follows the same trend as observed from 2011 to 2017, as short-term fluctuations mainly occur during statutory holidays, and the overall mobility trend remains relatively stable across years.

Other data

The data on Guangzhou’s Gross Domestic Product (GDP) and urban residents’ consumer price index for medical care and personal items are sourced from the annual statistical bulletins released by the Guangzhou Statistical Bureau. Detailed data are presented in Table S1.

Statistical analysis

Data preprocessing

Given the unique climatic and geographic circumstances of Guangzhou and the biannual influenza epidemic peaks in this region, we set the epidemic period for OTCCM% and ILI% as 49 weeks of the year to 48 weeks of the following year and chose 49 weeks of the current year to 16 weeks of the following year for the winter-spring epidemic season, and 17 to 48 weeks for the summer-fall epidemic season.

The OTCCM%, ILI%, ‘ILI exclude flu’, and meteorological variables were smoothed using a 4-week moving average to reduce short-term fluctuations and highlight underlying trends, following established practice in epidemiological and environmental time series analyses [21]. Regarding the identification of OTCCM% and ILI% epidemic peaks, we employed the simple average method to determine the ratio of weekly data to the annual average for each year. This ratio indicated a sales peak if it was more significant than 1. If there were more than three consecutive weeks where the percentage exceeded the yearly average, this was denoted as an epidemic peak [22].

Descriptive analysis

We conducted a descriptive analysis of the OTCCM%, ILI%, positivity rate for influenza subtypes, mean temperature, relative humidity, precipitation, sunshine duration, mobile population, in-migration index, and out-migration index in Guangzhou City from 2011 to 2017.

Wavelet analysis: We used Morlet wavelets, which have high-frequency resolution and can be localized on time scales, as the mother wavelet in this study. By decomposing the time series into time-frequency space, we could identify the primary variability patterns and observe how these patterns changed over time. The continuous wavelet transform was applied to individual time series, while wavelet coherence was used to assess the correlation between two time series in both time and frequency domains. We performed a continuous wavelet transform and wavelet coherence analysis on the time series of OTCCM% and the factors above to assess periodic patterns. This analysis used the R (version 4.2.2) package Rwave (version 2.6–5) and biwavelet (version 0.20.21).

Correlation analysis: We plotted a matrix of correlation coefficients for the various factors. The degree of correlation between two time series at different points in time was determined using the cross-correlation function (CCF). First, the correlation between the two time series of data was calculated using Spearman rank correlation analysis. Then, the one-time series was shifted forward and backward for the other time series. This study was performed in 1-week units. Considering reality and prior knowledge, the maximum unit length was restricted to 2 weeks.

Model construction

We constructed a Generalized Additive Model (GAM) with a Gaussian link function, using OTCCM% as the response variable. Continuous explanatory variables were modeled with smooth functions (s() in the mgcv package), and Dengue was included as a categorical variable. The covariates consisted of meteorological factors (mean temperature, relative humidity, precipitation, and sunshine duration), population movement factors (mobile population, in-migration index, and out-migration index), and infectious disease factors (‘ILI exclude flu’, outpatient visits, A/H1N1, A/H3N2, B, and dengue status). The lag periods for each variable (0–2 weeks) were determined through cross-correlation analysis with OTCCM%. Model selection focused on minimizing the Akaike Information Criterion (AIC). We examined multicollinearity using variance inflation factors (VIF) and found all variables retained in the final model had VIF values < 6, indicating acceptable levels of collinearity. Model diagnostics were conducted using the gam.check() function. This analysis was conducted using R (version 4.2.2) with the mgcv package (version 1.8–36). The full model specification and spline smooth plots are provided in the supplementary material.

Subsequently, an extreme gradient boosting tree (XGBoost) model, which theoretically should handle highly complex, nonlinear data better, was employed to capture the dynamics of OTCCM%, incorporating the 13 influencing factors mentioned earlier with appropriate lags as features. Built upon a gradient boosting framework, the XGBoost model utilizes decision trees as essential weak learners, iteratively forming a robust ensemble model that can capture intricate nonlinear relationships among features while exhibiting resilience to noise and outliers [23]. During the training of the XGBoost regression model, a grid search algorithm was implemented, and the optimal parameters were determined through five-fold cross-validation, evaluating metrics such as R-squared, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The best-performing model was applied to the entire dataset to scrutinize the relationship between features and OTCCM%.

Given the opaque nature of the machine learning model, including the difficulty of comprehending the connection between inputs and outputs, we applied the SHapley Additive exPlanations (SHAP) method to quantitatively assess the contribution of various features to OTCCM%, to display the effect curves, and to calculate the proportion of each feature’s contribution [24]. Based on solid mathematical (cooperative game theory), the SHAP framework is employed to conduct explanatory analysis on machine learning model outputs. The core idea is to regard all features (independent variables) as contributors to the output, generate SHAP values for each predictive sample, and allocate these to each feature of the predictive samples. SHAP is a linear additive model as follows.

2-2

In the  sample with  influences, where the  feature is , the predicted value of the machine learning model for this sample is , and the model baseline value is ,  is the SHAP value for . When > 0, it indicates that the feature improves the prediction value and contributes positively to , and vice versa for a negative contribution.

To visualize the nonlinear effect of each variable on the predicted OTCCM%, we applied Generalized Additive Models to fit the SHAP values of each feature. Specifically, we used the LinearGAM() function from the pyGAM package to fit a smoothed curve between each feature and its corresponding SHAP values across all observations [25]. This approach helps to intuitively display the marginal effect of each variable on the model’s output while maintaining interpretability. The shaded area around the fitted line represents the 95% confidence interval. This analysis was performed using the scikit-learn package (version 1.0.2), pyGAM (version 0.9.1), and shap (version 0.41.0) with Python version 3.7.

Results

Data description

The medication sales data originated from the enterprise sales information system of a pharmaceutical chain limited company with 36 pharmacies in Guangzhou City. They were regularly collected and organized by dedicated personnel. Regarding the number of medication sales, both OTC sales and OTCCM sales troughed during the annual Chinese New Year holiday, which to a certain extent reflected the changes in people’s behavior of purchasing medication during the Chinese New Year period (Fig. 1A). From January 2011 to December 2017, the median OTCCM% was 0.70%. The Interquartile Range (IQR) was 0.52% and 0.85% (Table S2). Regarding sales trends, OTCCM% exhibited annual cyclicality, typically peaking in winter-spring and summer of each year. OTCCM% exhibited an upward trend, followed by a downward trend, and then an upward trend from 2011 to 2017, with the highest percentage of sales occurring in 2012. From the summer of 2013 through the end of 2014, OTCCM% showed multiple patterns with multiple peaks and a lower percentage of sales.

Fig. 1.

Trends in Over-the-Counter Medication Sales and Related Factors in Guangzhou (2011–2017). A Sales of over-the-counter cold and cough medication. B Influenza-like illness percentage and positive test rates for influenza A/H1N1, A/H3N2, and B viruses. C Meteorological factors, including mean temperature, relative humidity, precipitaion, and sunshine duration. D Population mobility indicators, including mobile population, in-migration index, and out-migration index. The blue dotted line marks Chinese New Year, the green dotted line marks May Day, and the red dotted line marks China’s National Day. Blue shading indicates the winter-spring seasons in Guangzhou, while red shading indicates the summer season. Gray lines represent raw data, and darker lines indicate data smoothed over four weeks

We calculated ILI% and positive detection rates for different influenza subtypes based on data from the Guangzhou influenza sentinel surveillance hospitals from 2011 to 2017. ILI% exhibits an annual periodicity, with annual peaks occurring in spring and summer, and peaks spanning winter, spring, and summer in certain seasons (Fig. 1B). A severe dengue epidemic occurred in Guangzhou from June to December 2014. Dengue is associated with fever symptoms and aches and pains in the acute febrile period; in contrast, ILI% showed a second peak in the summer of 2014. A/H1N1 and B were predominantly prevalent in Guangzhou in winter and spring, with B becoming prevalent later than A/H1N1 viruses. A/H3N2 was prevalent in the spring and summer, and it was observed that the peak ILI% in the summer coincided with the peak of A/H3N2.

The changes in meteorological factors exhibited pronounced seasonality (Fig. 1C). The mean temperature in Guangzhou demonstrated regular seasonal fluctuations from 2011 to 2017, ranging between 6 and 30 degrees Celsius (Table S2). It peaked during the summer months of June and July and gradually decreased afterward. Relative humidity varied between 47% and 94%, with its lowest levels typically occurring at the end of each fall. Precipitation was notably more frequent in spring and summer, with a maximum weekly precipitation of 330 mm. The average sunshine duration during summer was 37 h per week, higher than the average of 23 h per week in the other seasons (t = 8.23, P < 0.05).

The mobile population (a long-term indicator) in Guangzhou City decreased slightly from 2011 to 2013. Then it continued to increase from 2013 to 2017 (Fig. 1D). We obtained the GaoDe migration index for the period from 2019 to 2021. We used only the 2019 migration index to analogize the 2011–2017 migration index data (short-term indicators), considering the impact of non-pharmacological interventions in 2020–2021. The average weekly in-migration index was 377 (IQR: 344, 409), and the average weekly out-migration index was 383 (IQR: 351, 408). Guangzhou’s in-migration and out-migration indices exhibit higher fluctuations during holiday periods, such as Chinese New Year, Qingming Festival, May Day, Labor Day, Dragon Boat Festival, Mid-Autumn Festival, and National Day. In contrast, they show a relatively stable level of fluctuations during non-holiday periods. The average in-migration index was significantly higher during holidays (486) than non-holidays (385, t = 2.90, P < 0.05), and the average out-migration index was also significantly higher during holidays (471) than non-holidays (384, t = 2.81, P < 0.05, Table S2).

Periodicity and correlation analysis of OTCCM% and related factors

Preliminary exploration of the data suggests that OTCCM% and various factors may exhibit periodic patterns. We further employ wavelet analysis to identify the periodicity of OTCCM% and multiple factors. Cross-correlation analysis is then used to assess the linear correlation between variables.

The results of the continuous wavelet transform indicated that OTCCM%, mean temperature, relative humidity, precipitation, sunshine duration, ILI, A/H1N1, A/H3N2, and B all exhibited annual periodicity. At the same time, A/H1N1, A/H3N2, and B also demonstrated six-month periodicity simultaneously (Fig. S1, Table S3). Wavelet coherence elucidated the strength of the association between two time series. Mean temperature, relative humidity, precipitation, sunshine duration, ILI, and A/H1N1 were generally associated with OTCCM% within the annual periodic band. A/H3N2 and B are strongly associated with OTCCM% within the 1–2 year periodic band (Fig. 2A, Table S3).

Fig. 2.

A Wavelet coherence between OTCCM% and other factors. The vertical axis in the figure represents the period scale (years). The white shaded area indicates the region affected by edge effects, and the black line represents the 5% significance level calculated based on 500 Markov bootstrapped series. B Matrix of correlation coefficients between OTCCM% and other factors. *0.05, **0.01, ***0.001

During the observation period, 11 peaks in OTCCM% were recorded, as shown in Table S4. The correlation coefficient between OTCCM% and ILI% throughout the study period was 0.22 (P < 0.05). During the peak seasons of winter-spring for OTCCM sales, OTCCM% exhibited the highest correlation with ILI% from the same period or one week prior (r: 0.69 ~ 0.87, P < 0.05). Conversely, during the peak sales periods of summer or spring-summer, the highest correlation with ILI% occurred within one week later (r = 0.46, P < 0.05). The number of weeks between OTCCM% peaks and optimal cross-correlation effects of ILI% tended to be either 0 or 1 week within the most extended incubation period for influenza.

The results of the correlation analysis indicate a significant negative correlation between mean temperature and sunshine duration, as well as OTCCM%, with r values of −0.28 (P < 0.001) and − 0.26 (P < 0.001), respectively. Among population movement factors, the mobile population showed a significant positive correlation with OTCCM% (r = 0.20, P < 0.001). For infectious disease factors, B showed the strongest positive correlation with OTCCM% (r = 0.45, P < 0.001), followed by A/H1N1 (r = 0.24, P < 0.001). Dengue was negatively correlated with OTCCM% (r = −0.36, P < 0.001). In contrast, ‘ILI exclude flu’ and outpatient visits showed weak and non-significant correlations with OTCCM% (r = 0.06, P > 0.05; r = −0.06, P > 0.05). Since only approximately 11% of ILI% cases tested positive for influenza on average, ‘ILI exclude flu’ primarily reflected other respiratory infections identified in sentinel hospitals.

Effects and contributions of influencing factors

The results of the above analysis suggested that multiple factors may be associated with OTCCM%, and the relationship between them may be nonlinear; we could construct nonlinear models or more complex machine learning models to explore the effect curves of multiple factors on OTCCM%. To more realistically represent the effects of each factor, we first conducted cross-correlation analyses to determine the number of weeks lagged for each factor relative to OTCCM%. Relative humidity with a two-week lag, in-migration index with a two-week lag, out-migration index with a one-week lag, outpatient visits with a two-week lag, and A/H3N2 with a two-week lag were included in the GAM and XGBoost regression models. Other factors were not time-series adjusted (Table S3). The construction process and detailed results of the GAM model are presented in the Supplementary material (Appendix).

For the constructed XGBoost regression model of OTCCM% with various factors, we utilized the SHAP interpretability framework to demonstrate the contributions of each factor. Regarding the contribution, the top three factors were mobile population, B, and mean temperature (Fig. 3A). Infectious disease factors contributed 38.15%; B and A/H1N1 epidemics had a significant positive contribution on OTCCM%, and dengue contributed negatively to OTCCM%. The characteristic mobile population, which responded to the long-term trend, and the in-migration and out-migration indices, which responded to the short-term trend, contributed 38.11%. Meteorological factors contributed 23.74%; high temperature significantly negatively contributed to OTCCM% (Fig. 3A and B).

Fig. 3.

SHAP analysis of features in the XGBoost model. A SHAP values for each feature in the XGBoost model. Red and blue colors represent high and low feature values, respectively, and a SHAP value greater than 0 indicates a positive contribution. B Contribution from each feature. C Individual feature SHAP values and fitted curves. Gray dots reflect SHAP values; blue curves are derived by fitting SHAP values using the GAM model, and the light blue shaded regions indicate 95% confidence intervals for the curves

To understand the complex relationship between OTCCM% and influencing factors, we fitted effect curves to observe the finer contributions (Fig. 3C). Regarding meteorological factors, the effect curve of mean temperature was reversed J-shaped, and the effect curves of relative humidity and sunshine duration were V-shaped. Low temperatures (below 22℃), low humidity (less than 70%), or high humidity (greater than 83%), as well as sunshine duration of less than 10 h or greater than 34 h per week, all contribute positively to OTCCM%. After the temperature exceeds 27℃, the negative contribution of high temperature to OTCCM% gradually weakens. Higher precipitation (exceeding 85 mm) has a negative contribution to OTCCM%.

A continuous positive contribution to OTCCM% is observed regarding population movement factors when the mobile population exceeds 5.26 million (Fig. 3C). An increase in the in-migration index has a positive impact on OTCCM%. Between 354 and 444, the out-migration index has a positive promoting effect on OTCCM%, but as it continues to increase, it exhibits a negative impact. Regarding infectious disease factors, the impact curve of outpatient visits shows a V-shaped pattern, indicating that as hospital outpatient visits increase, OTCCM sales at pharmacies decrease. However, a continuous increase in outpatient visits has a positive contribution to the OTCCM%. As the incidence of ‘ILI exclude flu,’ A/H1N1, A/H3N2, and B infections increases, OTCCM% also shows an upward trend. Similar effects of these three types of factors were also found in the GAM model (Fig. S2).

Discussion

In previous studies, single factors such as temperature, air pollutants, influenza-like illnesses, and gastrointestinal diseases have primarily focused on understanding medication sales, with few studies highlighting the role of complex, multifactorial influences. Unlike prior research, our study integrates three categories of factors—meteorology, infectious disease, and population movement—yielding a more comprehensive dataset. Utilizing machine learning algorithms alongside interpretable tools, we first quantified the contributions of multiple factors influencing OTCCM sales in the subtropical mega-city of Guangzhou. Additionally, we systematically characterized OTCCM sales trends, periodicity, and correlations with each of these factors. These results suggest that over-the-counter medication sales data can serve as a valuable supplementary indicator for understanding population-level health dynamics. Our findings revealed that infectious disease factors contributed most significantly to OTCCM% (38.15%), with the top three influencers being mobile population, influenza B, and mean temperature, corresponding to the categories of population movement, infectious disease, and meteorology, respectively. These results provide insights into local consumer behavior patterns, assist in the rational planning of pharmacy services, ensure the safety of drug quantities, and mitigate public health risks [26].

Influenza B significantly contributed to OTCCM sales among the factors related to infectious diseases. Similar to monitoring data from the Los Angeles area, our study suggests that influenza B may serve as a sensitive indicator of OTCCM sales in Guangzhou City [25]. Respiratory viral diseases exhibit seasonal patterns, as common cold viruses and influenza viruses co-circulate during the winter months—for instance, human respiratory syncytial virus peaks in winter [27]. Surveys conducted in the United States and Indonesia found that over 60% of people prefer OTC medications for treating mild respiratory illnesses, such as coughs and colds, and 30–40% choose to buy non-prescription drugs [28, 29]. We categorized ILI% into three influenza subtypes, and ‘ILI exclude flu.’ Since only a minority of individuals seek hospital treatment for upper respiratory infections, Influenza B may reflect the prevalence of other winter respiratory pathogens in the population. Symptomatic monitoring of ILI in outpatient clinics might not capture this aspect.

The low level of OTCCM% during the summer of 2014 may be related to the dengue fever outbreak in Guangzhou. Additionally, the ILI% exhibited a second peak in late September 2014. Following the dengue fever outbreak, Guangzhou implemented comprehensive community-based intervention measures, including collaboration among multiple government departments, mosquito control, public health education, and community participation. These measures enhanced hospital surveillance, increased residents’ health awareness through widespread mobilization, and, to some extent, led more residents with fever symptoms to seek medical treatment at hospitals instead of directly purchasing medication from pharmacies [17, 18]. The effect curve of outpatient visits to hospitals shows a V-shaped pattern. An increase in outpatient visits may lead to a decrease in OTCCM sales in pharmacies. However, when the intensity of common respiratory infectious diseases is high, hospital outpatient clinics may struggle to meet the demand for medical care for everyone. This leads to a further increase in the public’s health response to hospital prescriptions and medication purchases at pharmacies.

We found that meteorological factors contributed to 23.74% of OTCCM%, reflecting the health effects of environmental exposure at the level of medication sales. Low temperatures contribute positively to medication sales, negatively contribute to medication sales when the average temperature exceeds 22℃, and the negative contribution diminishes when the temperature exceeds 27℃. Low temperatures may lead to mucociliary clearance deficiency, impair type I interferon signal transduction, affect the host’s airway defense mechanisms, increase susceptibility to respiratory pathogens, and increase OTCCM sales [30]. Studies report an inverted J-shaped, U-shaped, or V-shaped relationship between temperature and respiratory disease incidence or mortality rates [10, 11, 31]. These studies have focused on various respiratory medications (nasal and throat preparations, medications for obstructive airway diseases, antihistamines). High temperatures may impair virus-specific adaptive immunity, potentially affecting cardiovascular events and indirectly influencing the prescription of medications related to the entire respiratory system [32]. We observed a V-shaped effect curve for relative humidity, similar to its impact on the prevalence of respiratory pathogens. Environmental temperature and humidity influence the viability and stability of respiratory viruses, including influenza, by altering the properties of viral surface proteins, lipid membranes, and the ratio of droplet nuclei. Similar conclusions have been drawn from animal transmission studies, chemical analysis studies, and studies on infectious disease transmission [27, 33, 34]. Previous research has often relied on morbidity, hospitalization, and mortality rates, which reflect only severe cases and lack sensitivity to minor environmental changes. This study on OTCCM sales confirms that medication sales can serve as an indicator of the health impact of environmental exposure on respiratory diseases [10].

We found that the peak in OTCCM% sales during the summer may be associated with the prevalence of A/H3N2 in Guangzhou. The summer incidence of A/H3N2 surpasses that of A/H1N1 and B, affecting populations of all age groups [35, 36]. Several time points with high prevalence rates of A/H3N2, even considering a 2-week lag, negatively influenced factors. This suggests both the unavoidable fluctuations in medication sales data and the potential that the peak incidence period of A/H3N2 in Guangzhou, typically occurring from April to July each year, aligns with times of high temperatures, increased precipitation, and extended sunshine duration. These influences may overshadow the positive impact of specific A/H3N2 samples in the model. Therefore, the high intensity of A/H3N2 summer prevalence may not translate into significant changes in population purchasing behavior regarding medication.

We found that during the winter-spring peak of OTCCM sales, the correlation between OTCCM% and ILI% from the same period or one week prior was highest, while during the summer peak of sales, the highest correlation was with ILI% from the same period or one week later. It reflects differences in public health responses during the winter-spring and summer seasons regarding purchasing OTCCM after experiencing ILI symptoms. Studies have shown that the body’s immune defense mechanisms display seasonal fluctuations. Cold weather conditions can weaken immune function, and low-humidity environments may dry out the respiratory mucosa, increasing individuals’ susceptibility to respiratory symptoms and discomfort [37, 38]. Conversely, people tend to show greater resilience to illness during the summer months [39]. Viral-challenge studies have shown that while the probability of viral infection was higher during longer day lengths or sunshine duration (summer), clinical illness was more likely to develop during shorter day lengths (winter), suggesting that endogenous seasonal rhythms may influence the progression from infection to illness in humans [40]. In high precipitation, people tend to spend more time indoors. The peak periods of OTCCM% and the optimal cross-correlation with ILI% often occur within 0 or 1 week, coinciding with the longest incubation period of influenza. This suggests the feasibility of using medication sales monitoring for influenza and other respiratory infectious diseases surveillance in Guangzhou City. These findings are based on a single city and must be compared with similar studies in other regions for further evaluation.

We present a novel perspective that has been previously underexplored: the long-term increase in mobile populations in megacities may correlate with the upward trend in OTCCM%, similar to how increases in urban population size and internal mobility have been associated with influenza epidemics in prior studies [41–43]. In megacities, the mobile population is driven by socio-economic development, and urbanization may indirectly influence the spread of infectious diseases such as influenza and hemorrhagic fever with renal syndrome [44, 45]. High migration rates often coincide with holidays, when increased interpersonal contact can enhance disease transmission, which may in turn indirectly influence medication sales. This suggests that public health officials should monitor population movement within their jurisdiction and prepare medication reserves in advance of holidays. Between 2011 and 2017, the mobile population in Guangzhou showed strong correlations with regional Gross Domestic Product (r = 0.905) and the Consumer Price Index for medical care and personal items (r = 0.767), both of which were statistically significant. While the mobile population had the highest SHAP contribution in our model, we lack sufficient evidence to assert a direct causal effect on medication sales. Instead, it may act as a proxy for broader sociodemographic influences. The intricate relationship between population movement and medication sales requires further research with more granular and prospective data [14].

Several limitations need to be addressed in this study. First, pharmaceutical sales data are susceptible to media reports, changes in pharmacy operating hours during holidays, fluctuations in pharmaceutical company stock prices, consumer “stockpiling” of over-the-counter medicines, the convenience of residents purchasing medications through medical insurance, and local variations in consumer behavior. These factors create challenges for data analysis and model fitting [9]. Second, the lack of comprehensive data on respiratory pathogen prevalence, daily population movement, and air pollution in Guangzhou during the study period may decrease model accuracy and exaggerate the estimated influence of the included variables. Future analyses should incorporate public health indicators (e.g., hospitalization and consultation rates) and environmental exposures (e.g., air pollutants), while exploring potential interactions between air pollution and meteorological factors [46]. Third, while machine learning models effectively captured complex associations, they did not establish causality. Future studies could incorporate causal inference techniques, such as structural equation modeling or quasi-experimental designs, to better clarify whether observed relationships reflect true causal effects or merely correlations, thereby enhancing the robustness and interpretability of findings. Furthermore, this study did not account for demographic factors such as age, gender, socioeconomic status, education, and healthcare access, which should be considered in future research to better capture consumer heterogeneity. With the rise of online over-the-counter medication purchases, improved surveillance of both online and offline sales is necessary to obtain a more comprehensive understanding of consumer behavior. Additionally, while we confirmed that all variables in the final model had a VIF below 6, residual multicollinearity may still impact the interpretation of variable importance by influencing how contributions are allocated across correlated predictors and should be addressed in future models.

Conclusions

This study evaluated the impacts of meteorological, infectious diseases, and population movement factors on OTCCM sales in a subtropical megacity. We utilized machine learning and interpretable models to reveal their complex relationships. These findings offer valuable insights into local consumer behavior, providing guidance for pharmacies and public health officials in planning services, managing medication supplies, and mitigating public health risks. Additionally, our results suggest that medication sales data may have potential for integration into epidemiological surveillance systems. With appropriate modeling, such data could facilitate the detection of early warning signals, the prediction of epidemic peaks, targeted medication distribution, and the identification of high-risk areas, significantly enhancing the ability to respond to epidemiological emergencies. However, due to the observational nature of our study and the non-causal modeling framework, further research is necessary to validate these applications and assess the feasibility of using medication sales data as a supplementary indicator in early warning systems for infectious disease outbreaks and pharmacy operations.

Abbreviations

OTCCM

Over-the-Counter Cold and Cough Medications

OTC

Over-the-Counter

ILI

Influenza-Like Illness

Guangzhou CDC

Guangzhou Center for Disease Control and Prevention

PR

Positivity Rate

CCF

Cross-Correlation Function

GAM

Generalized Additive Model

VIF

Variance Inflation Factor

AIC

Akaike Information Criterion

XGBoost

Extreme Gradient Boosting Tree

SHAP

SHapley Additive exPlanations

Authors’ contributions

XD designed the study. JC, DW, and BZ collected the data. JC and DW performed the analysis. XD, JC, DW, YC, HL, JZ, ZC, WH, GW, and XZ interpreted the data. JC and DW prepared the manuscript. XD, ZZ, JL, JC, and JZ edited the paper. All authors read and approved the final report. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Science and Technology Planning Project of Guangdong Province [grant number 2021B1212040017], the Major Program of Guangzhou National Laboratory [grant number GZNL2024A01002], the National Key Research and Development Program [grant number 2022YFC2303800], the Shenzhen Science and Technology Program [grant number ZDSYS20230626091203007], the Key-Area Research and Development Program of Guangdong Province [grant number 2022B1111020006], the Key Project of Medicine Discipline of Guangzhou [grant number 2021-2023-11], and the Medical Scientific Research Foundation of Guangdong Province [grant number B2023083].

Data availability

Medication sales data were obtained from a third party and are not publicly available. Infectious disease data were from the Guangzhou Center for Disease Control and Prevention and required application for disclosure. Meteorological and population movement data were sourced from publicly available government or commercial institutions’ websites. The code is available upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.