In this retrospective study, we analyzed deidentified clinical data from pediatric patients who were admitted to the Children’s Hospital Capital Institute of Pediatrics’ ED between January 1, 2021, and December 31, 2021. Our study aimed to evaluate the demographic characteristics, clinical presentations, and medical visit patterns of these PED visits. We used various computational models for predicting waiting times and discussed factors associated with increased arrivals and extended waiting times (Figure 1).

We identified all patients admitted to the Children’s Hospital Capital Institute of Pediatrics’ ED between January 1, 2021, and December 31, 2021, using the hospital’s information system. There were a total of 201,491 PED visits from 136,225 anonymous patients over approximately 12 months. We gathered comprehensive medical information about each patient’s PED journey. Essentially, we recorded key time points for each PED visit, including registration, doctor consultation or treatment, prescription filling, and departure. Upon arrival and registration, triage nurses assessed the patients’ clinical conditions, and the triaged patients then waited to see a physician. Throughout their hospital journey, some patients receive their medication and leave without revisiting, while others require additional tests for their initial diagnosis. These patients return to the waiting room for further clinical consultation or treatment after these tests. In this study, the waiting time to see a doctor was calculated from the moment of registration upon arrival to the first physician visit. Additional collected data encompassed patient gender, age, registration department, triage levels, diagnosis, and the number of doctors on duty.

We used 8 models to predict waiting times for PED visits. The rolling average estimator and linear regression (LR) were used as baseline models, while machine learning methods, including K-nearest neighbor (KNN), random forest (RF), LightGBM (light gradient-boosting machine), and XGBoost (extreme gradient boosting), were also used. The rolling average calculates short-term trends using a set of data. In short, with t_reg as the registration time stamp of a patient and n as a rolling time, the rolling average estimator predicts a patient’s waiting time by calculating the average waiting time of patients whose registration time stamps t′_reg and consultation or treatment time stamps t′_con fall within the interval t_reg – n, t_reg. In our study, we set n = {4h,2h} and constructed 2 rolling average estimators called Rol.Avg.4 h and Rol.Avg.2 h.

We used 10-fold cross-validation to evaluate the model performance. Additional information on the model hyperparameter search experiments conducted during this process is available in Multimedia Appendix 2. To mitigate random errors resulting from data splitting, we performed 10 repeated cross-validations with different random seeds and reported the average performance and SD for linear regression and machine learning methods. Additionally, we used Shapley additive explanation values, a successful tool in clinical predictions [27,28], to evaluate variable importance and select those most influential in waiting time prediction.

In our study, we used a Python script (Python Software Foundation) that used various imported application programming interfaces (scikit-learn=0.23.2, xgboost=1.2.0, lightgbm=3.0.0, shap=0.39.0) for model development. Model performance was assessed using R², mean absolute error, and root-mean-square error (RMSE).

All methods were performed in accordance with the relevant guidelines and regulations. As we used anonymized and deidentified data and did not constitute human research, the need for written informed consent was waived by the Ethics Committee of Children’s Hospital Capital Institute of Pediatrics due to the retrospective nature of the study (approval number SHERLLM2022034).

We collected data on a total of 183,024 eligible PED visits, involving 127,368 pediatric patients, during the study period. Of these, 56.90% (104,147/183,024) were male patients and 43.10% (78,877/183,024) were female patients (Table 1). The PED had an average of 501.44 visits/day (range 179-1145 visits/day). The majority of eligible admissions (175,791/183,024, 96.05%) were assigned to the green zone, indicating fewer complex conditions and less urgent treatment categories. Furthermore, we observed repeated visits to the PED, with 9.12% (11,612/127,368) of patients undergoing more than 2 emergency visits during the study period. Additionally, 18.76% (34,339/183,024) of PED visits resulted in discharge without a prescription or further tests.

The median waiting time from registration to seeing a doctor upon arrival was 27.53 minutes (IQR 8.48-83.25 minutes). Furthermore, 43.59% (79,786/183,024) waited less than 20 minutes, and 66.72% (122,120/183,024) waited less than 60 minutes. However, nearly one-fifth of patients waited over 100 minutes for clinical consultation or treatment. Comparatively, daytime PED visits had a median waiting time of 21.92 minutes, shorter than nighttime visits (35.90 minutes). Additionally, patients triaged in the yellow zone had a median waiting time of less than one-third of those triaged in the green zone (Table 1). Furthermore, pediatric patients registered in the emergency internal medicine department experienced longer waiting times compared with those in the emergency surgery department (median waiting time: 38.57 minutes versus 8.35 minutes, respectively; P<2.2×10^–16). This may be because the emergency surgery department often handles relatively simple procedures, such as dressing changes or minor wound and burn treatments. Pediatric patients with moderate to severe symptoms typically received emergency care in the internal medicine department, which often resulted in longer waiting times.

In this retrospective study, 71.26% (130,423/183,024) of PED visits were from children under 5 years of age, and 41.65% (76,228/183,024) were from those under the age of 3 years. The mean age of PED visits was 4.14 (SD 3.18) years. Patients under the age of 3 years had shorter waiting times compared with other groups (P<2.2×10^–16), with median waiting times of 25.80 and 28.82 minutes, respectively. Children with no more than 2 PED visits had a higher percentage of longer waits compared with those with 3 or more visits (median waiting time: 28.85 minutes versus 23.93 minutes). Fever (50,715/183,024, 27.71%), respiratory infection (43,269/183,024, 23.64%), celialgia (9560/183,024, 5.22%), and emesis (6898/183,024, 3.77%) were the leading causes of emergency room visits among children. Patients with respiratory tract infections often have fever symptoms. As depicted in Figure 2A, pediatric patients presenting with fever, respiratory infections, cough, pneumonia, bronchitis, or bacterial infections in the ED exhibited similar bimodal distributions across various age groups, specifically in the 1-2– and 3-4–year-old categories. When comparing the frequency of the top 20 leading causes of PED visits in 2 age groups (under 5 years old and 5 years old and older), several respiratory diseases (bronchitis, n=3227; pneumonia, n=2173; upper respiratory tract infection, n=1119; asthmatic bronchitis, n=1040; and laryngitis, n=926) were observed as common causes in the under 5-year-old age group, while more general symptoms (chest pain, n=626; headache, n=572; dizziness, n=432; and chest distress, n=411) occurred in the 5-year-old and older age group.

On average, there were 490.54 PED visits/day on weekdays and 528.78 visits/day on weekends. Compared with weekdays, the number of PED visits was lower, with Sunday having the highest count of PED arrivals, averaging 532.02 visits/day (Figure 2B). Similarly, we observed longer waiting times on Sundays compared with other days (Figure 2C). Regarding the temporal trends of daily admissions, on average, the number of PED visits was high between 10:00 AM and 11:00 AM and between 8:00 PM and 9:00 PM, while it was low between 2:00 AM and 5:00 AM (Figure 3A). In summary, there was a consistent increase in PED arrivals starting at 6:00 AM, reaching a peak at 10:00 AM on weekends and 11:00 AM on weekdays. Afterward, the number of PED visits started to decrease. Several hours later, it began to rise again, reaching a second peak around 8:00 PM, marking the busiest period in the emergency room.

We also noted the influence of seasonality and holidays on PED arrivals. Throughout the year, the ED admissions were at their lowest during early February (coinciding with the Spring Festival and winter holidays) and late August (coinciding with the summer holidays; Figure 3B). The number of PED visits was generally higher during academic semesters compared with breaks. In particular, the daily number of PED visits increased steadily throughout March and April, peaking in early May and mid-June, respectively (Figure 3B). We observed a positive correlation between the number of PED visits and the average waiting time. For instance, the waiting time for patients admitted in May and June was higher than that in February, indicating longer waits during these 2 months. Similarly, as the number of PED visits decreased in late August, the median waiting time also decreased and remained relatively low around September. Subsequently, the daily number of PED visits gradually increased as temperatures dropped during the winter months. The seasonality of common pediatric infectious diseases may have contributed to these fluctuations in PED admissions. Additionally, the proportion of PED visits among those aged 3-5 years decreased from late January to February (Figure 4). During the same time interval, there was an increasing trend in the proportion of pediatric patients younger than 3 years, similar to what was observed in late August and early September. By contrast, the number of pediatric patients aged over 5 years remained relatively stable.

We also examined the temporal variation in common causes of pediatric emergency room visits and their associated waiting times. As depicted in Figure 5A, patients with open or closed injuries typically experience shorter waiting times compared with those with inflammations (eg, pneumonia and gastroenteritis), infections (eg, respiratory infection), or general symptoms (eg, cough and emesis). This finding aligns with the medium waiting times observed in different EDs, as shown in Table 1. Some common causes of PED visits exhibited seasonal epidemics. For instance, fever, respiratory infection, gastroenteritis, and bronchitis had peaks in April to July and October to December, mirroring the overall pattern of PED visit changes shown in Figure 3B. In summary, emesis and cough peaked from April to July, while pneumonia was more prevalent from October to December. On the contrary, closed brain injury, open face injury, and open scalp injury showed stable occurrence patterns throughout the year (Figure 5B). To explore the relationships between shifting trends in common causes of PED visits and extended waiting times, we identified the top 5 common causes for each month (Figure 5C), as well as the causes with the top 5 monthly median waiting times (Figure 5D). Figure 5C demonstrates that the occurrences of fever and respiratory infections fluctuated substantially over the year, whereas other causes remained relatively stable. Figure 5D indicates that causes with higher waiting times exhibited similar shifting trends over the entire year, which also aligns with the overall average waiting time changes shown in Figure 3B.

The performance metrics for waiting time prediction models are presented in Table 2. Overall, the LR, KNN, RF, and gradient boosting machine models demonstrated higher accuracy and goodness of fit compared with the commonly used rolling average method. This is because the rolling average estimator tends to underestimate or overestimate the waiting time more than other methods. Among the machine learning models, RF, LightGBM, and XGBoost outperformed LR, with no significant difference in performance, and their combined average performance increased R² by approximately 29.33% (0.154/0.525) compared with LR and decreased RMSE by approximately 17.73% (8.951/50.481) in cross-validation. Machine learning models exhibited a broader range of values and had fewer predictions with high bias, as evidenced by the presence of more data points near the diagonal (refer to Multimedia Appendix 3). This aligns with existing studies demonstrating that machine learning models can capture more dynamic relationships between waiting time and related factors [22,25]. Among all the models, the XGBoost model achieved the highest R² and the lowest RMSE, consistent with previous studies [25].

Furthermore, because the predictors were encouraged to provide precise predictions, especially for patients with longer waiting times, it is important to note that these patients, although they constitute a small proportion of overall visits, have a substantial impact on the long waiting queues in the PED. Detecting and addressing these outliers are crucial for improved PED admission management. We computed the mean absolute error scores of waiting time prediction models for the 3 subgroups with relatively higher waiting times within the overall distribution (50%-75%, waiting time 27.43-82.92 minutes; 75%-95%, waiting time 82.92-209.34 minutes; and 95%-100%, waiting time 209.34-479.97 minutes). The results in Table 3 demonstrate that both LR and machine learning models outperformed the traditional rolling average methods, particularly in the extremely high waiting time subgroups (75-95% and 95%-100%). Conversely, LightGBM and XGBoost consistently maintained acceptable prediction performance across all abnormal subgroups.

Subsequent feature importance analysis revealed that the number of triage green patients waiting to see a doctor was the most important factor, followed by the type of registered ED (refer to Multimedia Appendix 4).

We compared the distribution of predicted waiting times from different models with the actual observed data [29]. Daily and monthly median waiting times, along with their 95th percentile intervals, were calculated for both the observed data and predictions from various methods (Figure 6). In Figure 6A, at the daily level, most machine learning models (RF, LightGBM, and XGBoost) demonstrated similar distributions of predicted waiting times compared with actual waiting times. By contrast, the rolling average methods did not consistently predict observations during crowding periods (eg, from May to July), and LR methods performed poorly during relatively noncrowded periods (eg, from January to May and from August to November). Similar trends were observed at the monthly level (Figure 6B), with machine learning methods (RF, LightGBM, and XGBoost) exhibiting slightly larger prediction deviations in the peak months of PED admissions. This could be attributed to machine learning methods tending to provide higher waiting time predictions when there are a substantial number of data points with extremely high values, leading to larger distribution deviations from the ground truth.

As the number of pediatric emergency visits increases, the demand for cost-effective emergency medical care will also grow. The emergence of big data and the widespread adoption of electronic medical records have enabled us to address population health issues once considered insurmountable. Previous studies have shown that big data and machine learning are emerging trends and essential tools for modern health care systems, capable of creating models that perform at human-level accuracy [30-32]. Using clinical records from 183,024 PED visits, this retrospective study examined the characteristics and admission preferences of pediatric emergency visits while predicting their waiting times in EDs. Unlike previous studies that concentrated on patients with specific injuries in the ED, our study encompassed nearly all pediatric patients admitted to the PED. This approach has potential implications for the comprehensive management and resource allocation of pediatric emergency medical care [33-35].

Our results indicated a median waiting time from registration to the first physician visit of 27.53 minutes, which was lower than that reported in previous studies [36,37]. However, as the demand for emergency services increased, we observed prolonged waiting times during specific periods. For instance, the number of emergency room arrivals and the median waiting time both increased from February to June, with fluctuations. Furthermore, a similar phenomenon was observed as temperatures dropped in winter. As shown in previous studies [38,39], common causes of PED visits, such as fever and respiratory infection in this study, also exhibited seasonal patterns (Figure 5B), which could influence PED admissions. Additionally, our data revealed that the majority of PED visits were from children under the age of 5 years (130,423/183,024, 71.26%). As mentioned earlier, an increased proportion of PED visits in children younger than 3 years was observed from January to February, while those aged between 3 and 5 years showed a decreasing trend (Figure 4). This is mainly because children under the age of 3 years are typically cared for at home, and they are known to have a higher incidence of respiratory tract infections during the winter. The dry and cold conditions during winter are major factors contributing to increased respiratory tract infections as they enhance virus stability and transmission while weakening the host immune system. By contrast, kindergartens typically begin their winter break in mid-January, reducing the likelihood of cross-infections among children aged 3-5 years. A similar phenomenon was observed in the summer holidays. Given that children are particularly vulnerable to diseases influenced by climate and school holidays, several measures could enhance the management of pediatric emergency medicine and reduce PED visits. These include creating extensive administrative databases to monitor seasonal disease patterns [40,41], identifying peak months of PED admissions for proactive physician and nurse practitioner scheduling in emergency care, and disseminating infectious disease prevention and management reminders to parents and school caregivers through mobile health (mHealth) apps [42]. However, these strategies depend on extensive data integration and collaboration among various teams in the EDs.

We also examined the daily peak times for emergency rooms and inferred potential reasons for admission preferences. As shown in Figures 2B and C and 3A, emergency rooms are typically busier on weekends, particularly on Saturday and Sunday mornings. Patients may prefer the emergency room on weekends, possibly because it is less convenient to schedule appointments with primary care physicians on Saturdays or Sundays. The pediatric emergency room experiences its peak activity between 10:00 AM and 11:00 AM and 8:00 PM and 9:00 PM, with fewer admissions between 2:00 AM and 5:00 AM. Interestingly, on weekdays, the emergency room tends to become significantly more crowded in the evening, except for children with urgent illnesses. This increase in crowding could be attributed to the higher influx of nonemergency patients, some of whom may be seeking care to avoid rush hours. Our data indicated that 18.74% (34,339/183,204) of PED visits resulted in discharge without a prescription or further tests. In effect, most outpatient services were unavailable after 5 PM, and those who failed to attend outpatient registration were more likely to choose an emergency room instead. This also results in overcrowded pediatric emergency rooms, indicating that these emergency rooms have limited capacity to deliver timely and quality care. The analysis of peak ED hours can inform policy decisions regarding medical resource allocation and treatment process optimization to reduce patient waiting times. For instance, the PED can schedule additional nurses and proficient emergency physicians during daily peak periods to ensure effective emergency care for children. Furthermore, increasing outpatient capacity, such as offering evening outpatient services to assess nonemergency PED admissions, could divert patients from the ED, alleviating congestion and reducing waiting times for pediatric patients in need of urgent care. Given the necessity of hospitalization for some urgent patients after their initial management in the PED, streamlining the process for transitioning patients from the ED to the inpatient setting can enable ED providers and nurses to serve a larger volume of patients and enhance the efficiency of clinical care [43]. This strategy could free up more treatment rooms for ED arrivals but relies on collaboration between the PED and other inpatient units within the hospital. As part of medical process optimization, the PED can implement a streamlined admission process during daily peak hours by combining registration and triage simultaneously. Nursing staff should receive training to rapidly and accurately assess whether a patient has a genuine emergency, and dedicated pediatric staff should be assigned to provide appropriate medical care for urgent patients [44]. Additionally, it may be necessary to use information models to enhance the triage of seemingly stable patients, monitor their short-term clinical changes, and offer timely treatment [45]. Other promising solutions are implementing online triage systems for parents or guardians to assess the urgency of their child’s symptoms [46,47], offering telehealth services for emergency care [48], and deploying social robots to provide emotional support to children during overcrowding periods [49]. For instance, Rochat et al [50] introduced a patient-centered mHealth app that encompasses all aspects of pediatric emergency care. This app offers symptom recommendations, predicts potential waiting times, enables patients to temporarily wait outside after triage using a queue reminder system, and notifies the ED upon the patient’s arrival. However, this was only tested in a laboratory environment, and further implementation and assessment in large-scale real-world data are necessary.

In this study, we also noted that children with fewer PED visits had a higher percentage of longer wait times compared with children with frequent visits. While all pediatric patients arriving at the emergency room require immediate care, those with frequent PED readmissions may have a higher likelihood of needing urgent medical attention. Further studies are needed to assess and compare the causes and timing of readmissions. Additionally, we conducted waiting time prediction and investigated the factors influencing waiting times. The most important factor was the number of triage green PED visits in the waiting room, followed by the registered ED. Further queueing theory analysis could be applied to emergency wards to reduce patient waiting times [51]. Machine learning methods outperformed traditional regression models in predicting waiting times. Implementing these prediction models in clinical practice can assist emergency staff in accurately assessing waiting times and managing crowded patient flow more effectively. In future research, the integration of additional data types should be considered. For instance, prior studies have demonstrated that integrating historical daily ED arrivals and internet search data improved daily ED volume prediction [52], which could directly impact waiting times. Simultaneously, waiting time prediction can aid health care workers in pursuing quality improvement initiatives to reduce patient wait times. As mentioned earlier, the number of triage green PED visits in the waiting room emerged as the most influential factor in waiting times, indicating a certain extent of medical resource shortage. The hospital implemented some optimization strategies, such as increasing the number of emergency physicians on duty at the end of June. As depicted in Figure 3B, this led to a decrease in waiting times for PED visits despite an increase in the number of PED arrivals. However, the impact of these policies was limited. Because of the high volume of PED arrivals, even a small increase in the number of doctors resulted in small effects. The shortage of pediatricians in China is alarming, and training an adequate number of skilled pediatric physicians and nurses quickly is challenging [53,54]. Therefore, it is essential to consider useful strategies to address this situation. This may include implementing broader policies that offer better benefits and working conditions for childcare workers and encouraging more medical students to pursue careers in pediatrics, particularly in PEDs. Recently, several key medical schools in China have reintroduced full-time majors in clinical pediatric medicine and related subjects. This initiative aims to boost the supply of pediatric medical resources and services [53].

Our study has several limitations. First, this was a retrospective study based on patient admission data from a single hospital. As a result, our findings may be influenced by regional health care patterns and epidemiological trends specific to the hospital’s location. This may limit the generalizability of our results to other pediatric health care systems. Second, we relied on standard features commonly used in previous waiting time evaluation studies. While these modeling strategies have been validated, there may be room for further optimization in their practical application. Additionally, some potentially relevant information might have been overlooked. Lastly, it is advisable to collect more data to confirm the seasonality of common pediatric infectious diseases and assess the impact of the pandemic on PED admissions, disease patterns, and care-seeking behaviors. Previous studies, primarily based on data from 2020 or 2019, have demonstrated that COVID-19 led to a decrease in admissions and ED visits at children’s hospitals [55,56]. The implementation of nonpharmaceutical interventions during the pandemic also resulted in a reduction in admissions for respiratory diseases among children in China [57]. In our future study, we plan to collect records over a longer time span. This extended time frame will allow us to capture detailed temporal trends in emergency admissions and enhance the robustness of our waiting time prediction model. In future research, we will consider exploring additional machine learning methods, such as deep neural networks, for waiting time prediction. Deep learning techniques, which involve artificial neural networks, can automatically learn features from data and potentially improve prediction performance and generalization [58]. Despite the limitations, our study highlights current trends and issues within the Chinese pediatric population seeking emergency care. These findings suggest potential quality-of-care challenges in pediatric emergency medicine and may inform policy decisions.

In conclusion, our real-world study analyzed the demographic characteristics, clinical presentations, and visit patterns of pediatric patients in the ED. We also developed machine learning models that effectively predict waiting times for doctor consultations. These models can aid physicians and patients in predicting busy periods and expected waiting times, facilitating timely decision-making for interventions. For children’s hospital EDs, they can be valuable in optimizing patient visit processes, significantly reducing waiting times, and enhancing overall pediatric emergency medical quality and service levels.