Comparison of SES method and SARIMA model in predicting the number of admissions in the department of neurology

Abstract

To establish and compare the prediction effect of SES and SARIMA model, and select the best prediction model to predict the number of patients in neurology department. The data came from HIS and medical record management system of a Grade-A hospital in Zhejiang Province. The number of inpatients from January 2019 to September 2023 was selected to establish SES and SARIMA model, respectively. Compare the fitting parameters, The larger the R²_adjusted, R², the smaller the RMSE, MAPE, MAE and standardized BIC, The better model is selected. Finally, the established model was used to predict the number of hospital admissions from October to December 2023, and the prediction effect of the MRE judgment model was compared. The number of admissions to the department of neurology shows a cyclical change, and drops sharply in January-February each year and rises rapidly in March. The best fitting models of SES model and SARIMA model were Winters addition model and SARIMA(0,1,1)(0,1,1)₁₂ model, respectively. The two models were selected to predict the number of admissions in the Department of neurology from October to December 2023, and the average relative error was 0.04 and 0.03, respectively. The prediction effect of SARIMA(0,1,1)(0,1,1)₁₂ model was better. Age and Spring Festival may be the factors that affect the periodic change of the number of admissions in neurology department. Both SES and SARIMA model can be used to predict the number of admissions in the department of neurology, and the SARIMA model may be better.

In recent years, with the development of society and medical treatment, the country’s requirements on the level of fine management of hospitals continue to increase, and the management mechanism of hospitals is facing new opportunities and challenges. Fine management has become the only way for the development of modern hospitals, and is a key factor to improve the operating efficiency and medical quality of hospitals. However, functional departments have rough management of clinical departments, and the setting of objectives is not scientific and rational, so the refinement level of hospital index management needs to be improved¹. Zhang et al.² showed that the current setting cycle of functional departments’ assessment goals for clinical departments was one year, and remained unchanged for 12 months. However, it is precisely due to the extensive management mode of the current hospital that the uniform management of the number of inpatients and the setting of the same target value throughout the year may lead to insufficient or waste of medical resource allocation. On the one hand, the actual number of inpatients exceeding the target number of inpatients will cause hospital congestion and further affect patient satisfaction, medical quality and treatment time, etc^3,4. On the other hand, the actual number of admissions less than the target number of admissions may cause a waste of medical resources. The accurate estimation of the number of hospital admissions is conducive to the reasonable arrangement and scientific allocation of medical resources, and can promote the efficiency of medical treatment and the fine management level of hospitals. Therefore, accurate estimates of admissions are particularly important.

The number of hospital admissions in different specialty departments and diseases in different months has seasonal changes, and the off-peak season is different⁵. A study in the United States found that a variety of diseases have obvious seasonal characteristics, such as the number of hospital admissions for various common pediatric diseases, including asthma, atrial septal defect, bronchiolitis, diabetic ketoacidosis, Kawasaki syndrome, and mental illness have obvious seasonal characteristics⁶. For the department of neurology, neurology mainly diagnoses and treats diseases related to the nervous system, mainly including cerebrovascular diseases, degenerative diseases of the nervous system, epilepsy and other seizure diseases, peripheral neuropathy and so on. A number of studies^7,8 have shown that common diseases in neurology show periodic changes. Cerebral infarction is one of the common cerebrovascular diseases in the department of neurology. The study results of Li et al.⁷ showed that the hospitalized patients with cerebral infarction showed an increasing trend year by year, and the number of hospitalized patients with cerebral infarction was the largest in the fourth quarter. Liu et al.⁸ also found that some neurological diseases have seasonal changes. For example, central nervous system infectious diseases are more likely to occur in warm seasons. In addition, the study found that in recent years, the number of patients hospitalized for intracranial infections and encephalopathy decreased significantly, but the number of patients hospitalized for central nervous system autoimmune diseases and hereditary metabolic encephalopathy increased year by year⁸. Therefore, we consider using this rule of cyclical change and seasonal fluctuation to predict the number of admissions to the department of neurology.

The prediction of the number of inpatients is a kind of prediction of time series data, which comprehensively considers the influence of various factors such as incidence, number of beds and policy factors⁹. The biggest feature of the time series is that the observed values are not independent, and the change law of the number of inpatients in the past is used to establish an appropriate model to predict the number of inpatients in the future. Predicting the number of hospital admissions is of great significance. Studies have shown that accurate estimation of the number of hospital admissions can determine the hospitalization rate in advance and promote resource allocation. For example, it can identify redundant idle beds and make reasonable arrangements¹⁰, thus providing a basis for optimizing the allocation of health resources⁹ and helping to improve the efficiency and quality of medical services¹¹.

At present, researches on hospital prediction mainly focus on disease mortality¹², disease development¹³, length of stay¹⁴, hospitalization cost¹⁵, hospitalization risk¹⁶, hospitalization outcome¹⁷, readmission of surgical patients¹⁸, nosocomial infection¹⁹ and other research fields. Aubert et al.¹² used HOSPITAL score and LACE index to predict the mortality rate of elderly patients with multiple diseases. Saengnipanthkul et al.¹³ Based on a multicenter prospective cohort study, the risk score of pediatric hospital-acquired malnutrition (PHaM) predicted the development of nutritional deterioration in hospitalized children. Wilk et al.¹⁴ used electronic records available at admission, including demographic, acute and long-term diagnosis, and physiological test results, to predict the length of stay of non-elective admission patients. Godara et al.¹⁵ found that in patients receiving allogeneic hematopoietic stem cell transplantation (allo-HSCT), patient and transplanting related characteristics as well as post-transplantation complications can be used to predict length of stay (LOS), which is also a driving factor for increased hospitalization costs. Leey-Echavarria et al.¹⁶ established a prediction model for the safety and hospitalization risk of reverse triage in the emergency department of the hospital. Weizman et al.¹⁷ predicted the hospitalization outcome of COVID-19 hospitalized patients with a scoring system based on machine learning. Based on data from 2,286 hospitals, Edington et al.¹⁸ explored whether quality indicators or hospital characteristics could predict readhospital penalties for hip and knee replacement. Tariq et al.¹⁹ predicted the risk of hospital acquired infection by using a fusion model based on graph convolutional networks.

However, most of the studies on the prediction of admissions focused on the whole country, the whole hospital, a certain department such as the emergency department^20,21,22, or a certain disease^23,24,25, and there was a lack of analysis on the number of admissions in neurology. During the COVID-19 pandemic in England, Meakin et al.²⁰ used three disease unknown prediction models, the average set of autoregressive time series models, and a linear regression model with local conditions with 7-day lag as predictors. And proportional convolution of local conditions and delay distributions for weekly projections of daily COVID-19 hospital admissions in UK National Health Service (NHS) trusts between August 2020 and April 2021, using weighted interval scores (WIS) to measure predictive performance. The results suggest that using confirmed COVID-19 cases as a predictor can improve admission prediction in some cases, but this is variable and depends on the ability to consistently make good case predictions. Wolff et al.²¹ used admission data from nine psychiatric hospitals in Germany from 2017 to 2020 to compare the model performance of the machine learning model with the weekly, monthly, and annual predictions of the time series model. Wang et al.²² used lasso regression to combine Google search information and COVID-19 related time series information with dynamic training and rolling window prediction to predict the number of new COVID-19 admissions in the United States at the national and state levels in the next two weeks. Bekker et al.²³ modeled the number of COVID-19 hospitalizations and occupancy in the Netherlands. Gong et al.²⁴ studied the current and future burden of hospital admissions for heat-related dementia in England. Culqui et al.²⁵ analyzed the relationship between environmental factors such as heat wave, noise and air pollutants and emergency admissions for Alzheimer’s disease in Madrid.

In addition, most of the existing literature focused on improving the accuracy of the prediction model, but did not specifically propose the actual resource allocation strategy (such as bed elastic management, etc.) matching the hospital management scenario^21,25,26. The main purpose of Wolff et al.²¹ ‘s study was to predict the number of admissions to psychiatric hospitals before and during the COVID-19 pandemic, as well as the performance of machine learning model and time series model. The best model in 2019 was machine learning model, and the best model in 2020 was time series model, without giving specific suggestions based on hospital scenarios. Culqui et al.²⁵ found that reducing PM 2.5 exposure in Alzheimer’s patients and providing special care to such patients during heat waves may lead to a reduction in emergency Alzheimer’s disease admissions and related health care costs, but did not provide a practical solution strategy for policymakers. Yang et al.²⁶ evaluated which model of the exponential smoothing prediction model and the SARIMA model could better fit the number of admissions, but gave few specific practical suggestions.

Both exponential smoothing method and SARIMA model belong to time series models and are often applied to the prediction of time series data^27,28. Therefore, this study intends to use exponential smoothing method and SARIMA model, two common time series models, to predict the number of patients admitted to the neurology department of a Grade A tertiary hospital in Zhejiang Province, and based on the research results, relevant strategies can be proposed for the reasonable allocation of medical resources by the hospital management department.

Materials and methods

Data source and research design

Data source

The data in this study came from HIS System and medical record management system of a grade A tertiary hospital in Zhejiang Province. This study complies with the guiding principles of the Declaration of Helsinki and was approved by the Ethics Committee of Zhejiang Provincial People’s Hospital. All patients gave informed consent.

Research design

This study was a retrospective time series analysis, and a prediction model was constructed based on historical data. In order to study whether the number of inpatients in the Department of Neurology of a grade A tertiary hospital in Zhejiang Province shows periodic regular changes, and to explore the reasons that may affect the change in the number of inpatients, we selected the monthly admission data of the hospital from January 2019 to September 2023 to establish a model. The monthly admission data from October to December 2023 were used to verify the fitting effect of the model, and the data were real and reliable. The output variable of this study was the number of admissions of the department of Neurology of a grade A tertiary hospital in Zhejiang Province from October to December 2023. For the determination of sample size, first of all, a number of studies reported 4–5 complete seasonal cycles to capture the seasonal pattern initially, satisfying the stability of the model and the accuracy of parameter estimation^29,30,31,32. Therefore, this study uses monthly data from January 2019 to September 2023 to build a forecast model with annual seasonal cycles (12 months), which actually contains 4.75 seasonal cycles for a total of 57 months of observations, enough to capture seasonal fluctuations and long-term trends. And this choice is also limited by the time frame of medical data opening (the data before the upgrade of the hospital information system is not available). Adopt the latest and complete data available in the hospital’s HIS system to avoid the deviation caused by artificial data acquisition. In addition, we also consider that in a rapidly changing healthcare environment, using data from the last five years can better reflect current care patterns than using data from a longer period, avoiding the introduction of bias due to outdated historical data. In this study, participant blindness and data collection blindness were used for the time series prediction model (SES or SARIMA), i.e. neither the participant nor the data collector knew the grouping information of the training set and the test set, which helped to reduce the possible bias of both the participant and data collector. The data analyst knows the grouping information, divides the participants into a training set and a test set based on the time of admission, and builds a predictive model. In addition, this study follows TRIPOD (Transparent Reporting of a multivariable prediction model)³³ and RECORD (REporting of studies Conducted using Observational Routinely-collected Data)³⁴ statements, where the TRIPOD statement is used to standardize the development and validation reports of prediction models (e.g., SES, SARIMA models), RECORD statements are used to standardize the reporting of observational studies based on routine medical data, such as hospital HIS systems.

Research methods

Original sequence diagram

The original sequence map was drawn for the number of admissions in the Department of Neurology from January 2019 to December 2023, and the change trend and periodicity were preliminarily judged according to the original sequence map.

Model introduction

Exponential smoothing prediction model

The exponential smoothing method is one of the moving average methods, which is characterized by giving different weights to the past observations, that is, the weights of the recent observations are larger than the weights of the future observations. The basic idea is that the predicted values are the weighted sum of the previous observations, and different weights are given to different data, the new data is given a larger weight, and the old data is given a smaller weight.

SES model is a seasonal extension of the exponential smoothing model. The exponential smoothing model is used to smooth and predict the time series, while the SES model takes seasonality into account on top of this.

The basic formula of exponential smoothing method is S_t=a*y_t+(1-a)* S_t−1, where S_t is the smoothing value of time t; y_t is the actual value of time t; S_t−1 is the smoothing value of time t-1; a is the smoothing constant whose value range is [0,1]. When the exponential smoothing method is used for modeling, the main steps are as follows:

(1) model establishment

The number of admissions to the department of Neurology was converted into time series data in the format of “year/month”, the information characteristics were displayed by time series graphs, and an exponential smoothing model was created.

(2) model selection

According to the higher R²_adjusted and R², RMSE, MAPE, the smaller the values of MAE and Standardized BIC, the better exponential smoothing prediction model is selected based on the principle of better fitting effect. Ljung-Box Q test or residual graph is used to determine whether the residual sequence is white noise. When P > 0.1 or the residual sequence values in the residual autocorrella and partial corrella are both within the confidence interval, it indicates that the residual sequence is classified as white noise sequence, indicating that the model has been able to explain the randomness of the data well.

(3) Model evaluation: The statistics of the model are analyzed. The exponential smoothing method involves three parameters: α (horizontal smoothing coefficient), γ (trend smoothing coefficient) and δ(seasonal coefficient). The values of α, γ and δ are all between 0 and 1, and when the parameter is close to 0, the influence weight of recent observations is smaller; The closer it is to 1, the greater its weight in the prediction^6,19.

(4) model prediction

Mean Relative Error (MRE) was used to evaluate the prediction effect of the model. MRE=$\frac{1}{n}\sum\limits_{{i=1}}^{n} {\left| {\frac{{yi - \mathop {yi}\limits^{{\text{\varvec{\Lambda}}}} }}{{yi}}} \right|}$, MRE is the average value of the predicted value and the actual value. MRE∈[0,1], a smaller value indicates a higher accuracy of the prediction ($\:{\widehat{y}}_{i}$ is the predicted value, y_i is the true value, the following formulas are the same).

SARIMA model

SARIMA model is a seasonal autoregressive moving average model, referred to as seasonal ARIMA, that is, adding a seasonal part on the basis of ARIMA. The SARIMA model can be expressed as SARIMA (p, d, q) x (P, D, Q) s, which satisfies the multiplication principle, the first part represents non-seasonal part, the last part represents seasonal part, s represents seasonal frequency [p: Autoregressive Order, d: Differencing Order (Differencing Order), q: Moving Average Order, P: Seasonal Autoregressive Order, D: Seasonal Differencing Order, Q: Seasonal Moving Average Order, s: Seasonal Differencing order Seasonal Period (Seasonal Period) is one of the predictive analysis methods of time series. When modeling with the SARIMA model, the main steps are as follows:

(1) sequence stabilization

Observe the original sequence diagram of the number of admissions in the department of Neurology to determine whether it is stable. If it is not stable, it needs to be stabilized, and differential and seasonal differential can be carried out until it is stable.

(2) model identification

The order and parameters of the model are determined according to the difference order and the autocorrelation function (ACF) graph and partial autocorrelation function (PACF) graph of the differential stationary sequence, and the SARIMA model is established accordingly. The optimal model is selected according to parameters R²_adjusted, R², RMSE, MAPE, MAE and standardized BIC values.

(3) model diagnosis

The significance test of model parameters is used to test the validity of the model. If P < 0.1, it means that the statistic passes the significance test. Ljung-Box Q test was used to test the residual of the model for white noise.

(4) model prediction

The model after fitting is used to predict future data, and the MRE is used to evaluate the prediction effect of the model. The Q-Q Plot (Quantile-Quantile Plot) and Shapiro-Wilk test are used to diagnose the normality of the model error. If the residual weight of the Q-Q plot is close to the reference line and there is no obvious systematic deviation, or the P-value of Shapiro-Wilk test is greater than 0.1, It indicates that the residual is normal.

Sensitivity analysis

For the exponential smoothing model, sensitivity analysis is performed by changing the parameters: α (horizontal smoothing coefficient), γ (trend smoothing coefficient) and δ(seasonal coefficient). Parameters d and D in SARIMA(p, d, q) (P, D, Q)s model have been determined according to data characteristics, so parameters p, q, P, Q are changed for sensitivity analysis. The stability of performance indicators such as R²_adjusted, R², RMSE, MAPE, MAE and standardized BIC was evaluated.

Cross-validation analysis

Due to the natural time dependence of time series data, temporal segmentation and Rolling time window cross-validation were used to evaluate the generalization of the model. In addition, since d = 1 and D = 1, other model parameters were taken into account and the period was 12, in order to fully capture the complete cyclical change of the number of admissions, 36 months was selected as the window length for the first time, and the training window was gradually expanded to predict the number of admissions in the next 3 months.

Window 1: January 2019 - December 2021 → Forecast January 2022 - March 2022.

Window 2: January 2019 - March 2022 → Forecast April 2022 - June 2022, and so on, repeating until traversing the entire data set.

In addition, we calculated prediction errors using multiple assessment measures (R²_adjusted, R², RMSE, MAPE, MAE, and standardized BIC).

Relevant formulas and definitions

R² = 1-$\frac{{\sum {(yi - \mathop {yi\mathop )\nolimits^{2} }\limits^{ \wedge } } }}{{\sum {(yi - \mathop {yi\mathop )\nolimits^{2} }\limits^{---} } }}$, R² represents the proportion of data variance that the model can explain; the numerator represents the sum of squared variance of the true value and the predicted value; the denominator represents the sum of squared variance of the true value and the mean value; R²∈[0,1], the closer R² is to 1, the better the model fits ($\overline{yi}$ representing the average value of the true value; the following formulas are the same).

R²_adjusted = 1-$\frac{{(1 - \mathop R\nolimits^{2} {\text{)(n-1)}}}}{{{\text{n-p-1}}}}$, Where n is the number of samples and p is the number of features. R²_adjusted∈[0,1], Adjusted R-Square offset the influence of sample size on R², the greater the stable R², the better the model fitting.

RMSE=$\sqrt {\frac{1}{n}\sum\limits_{{i=1}}^{n} {\mathop {(yi}\limits^{ \wedge } } - \mathop {yi)}\nolimits^{2} }$, RMSE is the square root of the sum of squares of the difference between the predicted value and the true value. RMSE∈[0, +∞). The smaller the RMSE value is, the smaller the prediction error of the model and the stronger the prediction ability of the model.

MAPE=$\sqrt {\frac{1}{n}\sum\limits_{{i=1}}^{n} {\mathop {(yi}\limits^{ \wedge } } - \mathop {yi)}\nolimits^{2} }$, MAPE represents the average percentage of relative error between the predicted value and the actual value. MAPE∈[0, +∞), when the real value has data equal to 0, there is a problem of dividing the denominator 0, $\frac{1}{{\text{n}}}\sum\limits_{{i=1}}^{n} | \mathop {yi}\limits^{ \wedge } - yi|$ this formula is not available.

MAE=$\frac{1}{{\text{n}}}\sum\limits_{{i=1}}^{n} | \mathop {yi}\limits^{ \wedge } - yi|$, MAE is the average absolute error between the predicted value and the true value, MAE∈[0, +∞), the smaller the MAE, the smaller the error.

BIC = ln(n)k–2ln(L), BIC is derived from Bayesian inference, which is to estimate the partial unknown state with subjective probability under incomplete information, and then modify the occurrence probability with Bayesian formula. Finally, the optimal decision is made using the expected value and the modified probability to balance the goodness of fit and complexity of the model, so as to select the best model. Where, k is the number of model parameters, n is the number of samples, and L is the likelihood function. Smaller BIC indicates that the model achieves a better balance between fitting data and model complexity, and the higher the model fit degree.

Statistical analysis

Excel 2016 was used to organize the data, and the missing values were filled with multiple interpolation method. The outliers with obvious errors in the input data are eliminated. Quantitative variables were represented by $\overline X$ ±S and qualitative variables were represented by n (%). The Analyze-Forecasting module of SPSS 22.0 software was used to establish and verify the SES model and SARIMA model. Spearman correlation analysis was used to explore the correlation between age, length of stay, Spring festival factor, and rate of complications and the number of admissions to the department of Neurology. Statistically significant variables in correlation analysis were incorporated into the prediction model as control variables to establish a prediction model. In this study, the model parameters of the SARIMA model were tested by t-test, and the null hypothesis (H₀) was that the model parameters were equal to zero. The alternative hypothesis (H₁) is that the model parameters are not equal to zero. It is used to evaluate whether each coefficient in the model (such as p: Autoregressive Order, q: Moving Average Order) is significantly different from zero, that is, to judge whether the prediction of these parameters on the time series is statistically significant. If the p-value of a parameter<0.1, it indicated that the parameter had a significant contribution to the model and should be retained in the model. If the p-value > 0.1, it may indicate that the parameter is redundant and try to simplify the model. The Ljung-Box Q test was used to test the hypothesis of the residual series of the model. The null hypothesis (H₀) is that the residual sequence is a white noise sequence. The alternative hypothesis (H₁) is that the residual sequence is not a white noise sequence. If p-value > 0.1, the null hypothesis cannot be rejected, and the residual sequence of the model cannot be rejected is a white noise sequence, indicating that the model information has been extracted by the model and the model is appropriate. Otherwise, the model should be readjust. Test level α = 0.1.

Results

Original sequence diagram

From January 2019 to December 2023, the total number of patients admitted to the Department of Neurology in a Grade-three hospital in Zhejiang Province was 24,409. The monthly number of patients admitted to the department of neurology during this period was sorted out and the sequence diagram was drawn. It was found that the number of patients admitted to the department of neurology dropped sharply from January to February each year and rose rapidly in March, showing obvious cyclical changes, as shown in Table 1; Fig. 1.

Table 1 Number of admissions in the department of neurology from January 2019 to December 2023.

Fig. 1

Sequence chart of admissions in the Department of Neurology from January 2019 to December 2023.

Model fitting and forecast

Exponential smoothing model

Model Building and selection

In this study, three SES models, namely simple seasonality model, Winters addition model and Winters multiplication model, were constructed, and the fitting parameters of each model were shown in Table 2. By comparing the fitting parameters of the models and combining with the selection principle of evaluation indicators, the Winters addition model in the exponential smoothing model was finally determined to be the best model (R²_adjusted = 0.461, R² = 0.863, RMSE = 40.203, MAE = 29.402, standardized BIC = 7.601).

Table 2 Model fitting parameters of different seasonal exponential smoothing models.

According to the results of Ljung-Box Q test, P > 0.05, indicating that there is no autocorrelation and partial autocorrelation in the residual sequence after data fitting, and this model can be used for prediction (Table 3). Besides, according to the residual autocorrelation and partial correlation, the values of the residual sequence fall into the confidence interval, indicating that there is no autocorrelation and partial autocorrelation in the model residual sequence, and it is a white noise sequence. See Fig. 2.

Table 3 Ljung-Box Q test results of different exponential smoothing models.

Fig. 2

Residual diagram of Winters addition model.

The statistical analysis results of different exponential smoothing models showed that the three models had statistical significance in α (level), and the Winters multiplication model had statistical significance in δ(season) (P < 0.1). However, the simple seasonal model and the Winters addition model had no statistical significance in δ(season), and the Winters addition model and the Winters multiplication model γ(trend) had no statistical significance (P > 0.1), suggesting that the number of admissions in the department of neurology may be seasonal. However, the simple seasonal model and Winters addition model showed that the decomposed seasonal characteristics were not obvious. See Table 4.

Table 4 Statistical analysis of different exponential smoothing models.

Model prediction

We used the Winters addition model to predict admissions from October to December 2023. The results showed a MRE of 0.04, and all admissions predictions in the Department of Neurology fell within 95%CI, indicating good prediction results. The results are shown in Table 5; Fig. 3.

Table 5 Winters addition model predicts admissions in the department of neurology in October-December 2023.

Fig. 3

Prediction of admissions in the Department of Neurology in October-December 2023 with Winters’ addition model.

SARIMA model

Sequence stabilization

The original sequence diagram shows that the sequence is unstable and periodic. The sequence diagram after trend difference (d = 1) and periodic difference (D = 1) can be seen to be basically stable, See Fig. 4.

Fig. 4

Sequence diagram of the number of admissions in neurology department after first-order trend difference and first-order periodic difference.

Model recognition

The SARIMA (p, 1, q) (P, 1, Q) ₁₂ model is established according to the result of the stabilization, that is, the difference order d = 1, D = 1. The ACF plot and PACF plot of the sequence after difference show the truncation of ACF diagram 1 and PACF diagram 1, so q may take 0 and 1, and p may take 0 and 1. Besides, considering the seasonal autocorrelation characteristics of the series, the ACF plot after difference shows that the delayed 12th-order autocorrelation coefficient is significantly zero, so it can be judged that Q may be 0 or 1. The PACF plot after difference shows that the partial autocorrelation coefficients of order 12 delay are significantly zero, so P may be 0 or 1. See Fig. 5.

Fig. 5

ACF plot and PACF plot of the sequence after difference.

Fit all possible reasonable models, compare the fitting parameters of different models, and select models whose model parameters make sense. The results showed that SARIMA (0,1,1) (0,1,1) ₁₂ model, SARIMA (0,1,1) (0,1,0) ₁₂ model, SARIMA (0, 1, 0) (0, 1, 0) ₁₂ model, All parameters of the 3 models are meaningful, and SARIMA (0, 1, 1) (0, 1, 1) ₁₂ is the best model (R²_adjusted = 0.239, R² = 0.771, MAE = 39.161). See Table 6. The cross-validation results showed that the R²_adjusted = 0.381, R² = 0.668, RMSE = 50.785, MAPE = 11.953, MAE = 36.338, standardized BIC = 8.481, SARIMA (0,1,1) 12 model, RMSE = 50.785, MAPE = 11.953, MAE = 36.338. There is no significant difference between the model evaluation index and the training set, which indicates that the model has good generalization.

Table 6 Model fitting parameters of different SARIMA models.

Spearman correlation analysis showed that age, length of stay and Spring Festival were correlated with the number of admissions in the department of Neurology (P < 0.1). Therefore, we included these three variables as control variables to establish the SARIMA model. See Table 7.

Table 7 Correlation between the number of admissions in neurology department and independent variables.

SARIMA (0,1,1) (0,1,1) ₁₂ model parameter test results showed that MA(1), MA(1), seasonal, average age and Spring Festival factor t statistics passed the significance test (P < 0.1), indicating that there was a significant cyclical change in the number of admissions. The factors of age and Spring Festival were the factors influencing the periodic change of the number of admissions in neurology department. See Table 8.

Table 8 SARIMA(0,1,1)(0,1,1)₁₂ model parameter tests.

Model diagnosis

Ljung-Box Q test was used to test the residual of SARIMA (0,1,1) (0,1,1) ₁₂ model for white noise. The results showed that the test statistic was 6.979, and the difference was not statistically significant (P = 0.958), indicating that the residual sequence was classified as white noise sequence, that is, the residual sequence was classified as pure random sequence. It shows that the SARIMA (0,1,1) (0,1,1) ₁₂ model is suitable. In addition, according to the residual ACF and PACF plots, the values of the residual sequence fall into the confidence interval, indicating that the established model performs well. See Fig. 6. In this study, the Q-Q plot was used to diagnose the normality of the model error. The results showed that the residual handicap was close to the reference line without obvious systematic deviation, indicating that the residual error was normal. See Fig. 7. In addition, Shapiro-Wilk test results showed that the model residuals followed a normal distribution (W = 0.989, P = 0.956).

Fig. 6

SARIMA (0,1,1) (0,1,1) ₁₂ model residual ACF plot and PACF plot.

Fig. 7

Q-Q plot of residual normality of SARIMA (0,1,1) (0,1,1) ₁₂ model.

Model prediction

SARIMA (0,1,1) (0,1,1) ₁₂ model was used to predict the number of admissions in the department of neurology from October to December 2023. The results showed that the mean relative error was 0.03, and all the predicted values of the number of admissions in the department of neurology fell within 95%CI, indicating that the prediction results performed well. See Table 9; Fig. 8.

Table 9 SARIMA (0,1,1) (0,1,1) ₁₂ model predicts admissions to the department of neurology in October-December 2023.

Fig. 8

SARIMA (0,1,1) (0,1,1) ₁₂ model predicted the number of admissions to the Department of Neurology in October-December 2023.

Sensitivity analysis

The change of parameter sum in the exponential smoothing model and SARIMA model has little influence on the prediction effect, and the results are more robust. See Tables 2 and 6.

For the exponential smoothing model, the Winters addition model is more adaptable to data characteristics by separating trend and seasonal components (addition overlay). The Winters multiplication model (multiplying trend with seasonality) may amplify noise interference. When the smoothing coefficient is selected, α (level), γ (trend) and δ (season) are well balanced in the Winters addition model, avoiding overfitting (low standardized BIC). The exponential smoothing model is sensitive to the seasonal parameter (δ), and the effectiveness of seasonal adjustment directly affects the performance of the model. Winters addition model is preferred, which has lower parameter sensitivity and higher stability.

For the SARIMA model, increasing p or q increases R² slightly (e.g., R² rises to 0.785 when p = 1), but the standardized BIC rises simultaneously. Most models with q = 1 (such as SARIMA(0,1,1)(0,1,1)₁₂) have lower RMSE and MAE, indicating that the moving average term is effective for noise suppression. Seasonal parameters (P = 1 or Q = 1) had a limited improvement in MAPE (SARIMA(0,1,1)(0,1,1)₁₂ had a MAPE of 11.252, slightly better than a model without seasonal adjustment). SARIMA(0,1,1)(0,1,1)₁₂ has moderate sensitivity and best predictive stability.

Comparison of model prediction effect

In the SES model, the Winters addition model performed well, while in the SARIMA model, the SARIMA (0,1,1) (0,1,1) ₁₂ model performed well. The MRE of the two models were 0.04 and 0.03, respectively, to predict the number of hospital admissions from October to December 2024. The prediction effect of SARIMA (0,1,1) (0,1,1) ₁₂ model is better.

Discussion

In our study, the SARIMA (0,1,1) (0,1,1) ₁₂ model was the most effective in predicting the number of new admissions in the neurology department of a top-three hospital in Zhejiang Province.

We draw the original sequence diagram, the number of neurology hospital admissions found every 1–2 month significantly lower, rapid rise in March, volatility change appears in other, similar to the related research results^35,36, Chiang, etc³⁵. the study found that China’s Taiwan in January and February hospitalized per day are significantly higher than other month, Chen et al.³⁶ also found that the number of patients in the neurology department of a Grade III and A class A hospital in Chongqing was low in January and February, which may be due to the Spring Festival holiday. Chinese people traditionally believe that the Spring Festival is a time for reunion. Chinese people are more likely to go to the hospital before or after the Spring Festival, rather than during the Spring Festival. Therefore, the Spring Festival factor was also considered as one of the predictors of the number of admissions in the department of neurology when the prediction model was established in this study, which also provided a reference for the hospital management department to rationally arrange medical resources. For hospital management, first of all, the dynamic adjustment of hospital human resources can be made. For example, due to the significant decline in the number of hospital admissions during the Spring Festival low period (January-February), the scheduling of medical staff in non-emergency departments can be reduced, and some staff can be deployed to emergency departments or other departments with stable demand to avoid idle human resources. After March, it gradually enters the peak period, and the allocation of medical personnel in neurology and related departments (such as imaging department and rehabilitation department) can be increased in advance to ensure timely admission of patients and shorten the waiting time. Then, the beds can be flexibly managed. Before and after the Spring Festival, the idle beds in the neurology department can be temporarily allocated to other departments (such as respiratory department and cardiovascular department) to improve the bed utilization rate. And recover and reserve enough beds before the peak. Finally, hospital managers can optimize equipment and drug stocks by adjusting stocks of drugs (such as stroke emergency medications) and examination equipment (such as MRI and CT) based on seasonal forecasts to ensure adequate supply during peak periods and avoid delays in treatment. This study also provides scientific basis for policy making. First of all, regional medical resources coordination. Based on the seasonal forecasting model, government departments can gradually allocate resources (such as financial subsidies and temporary medical team support) to hospitals with a concentration of neurology patients after the Spring Festival to ease regional medical pressure. Second, strengthen public health education strategies. In response to the phenomenon of “delayed medical treatment” during the Spring Festival, health education activities are carried out to emphasize the importance of early medical treatment for non-emergency symptoms (such as dizziness and numbness of limbs), and reduce the medical congestion caused by centralized medical treatment after the holiday.

In addition, considering that the number of admissions to the Department of Neurology in this study showed seasonal cyclical changes, we added seasonal factors in the establishment of the model, established the SES model and SARIMA model. In this study, the traditional time series approach (SES/SARIMA) was chosen over the machine learning model based on the following considerations: Interpretability requirements: Clinical managers prefer intelligible models to support decision making (such as seasonal decomposition of SARIMA); Small sample fit: The size of the original dataset is small (n = 57), and the complex model is easy to overfit. And machine learning (such as LSTM, Random Forest, or Gradient Boosting) also has some challenges: such as black box problems, over-reliance on data volume (small samples are easy to overfit). In addition, the focus of this study is to verify the applicability of the traditional time series model (SES/SARIMA) in the prediction of the number of admissions in the department of neurology, aiming to provide a reference for the optimization of medical resource allocation in medical institutions with limited resources, and the traditional time series method can fully meet the requirements. In the future, we look forward to building models with a hybrid “traditional model + machine learning” framework for a more comprehensive evaluation of healthcare forecasting methods.

In this study, SARIMA model was selected to predict the number of admissions in neurology department, which is a model with strong Interpretability. Firstly, we can evaluate the interpretability of the model by parameter significance test, and determine whether the coefficients of AR (autoregressive) and MA (moving average) terms are significant by T-test or P-value (p < 0.1). In this study, SARIMA (0,1,1) (0,1,1) ₁₂ model parameter test results showed that MA(1), MA(1), seasonal, average age and Spring Festival factor t statistics passed the significance test (P < 0.1), indicating that there were obvious periodic changes in the number of admissions. The factors of age and Spring Festival were the factors influencing the periodic change of the number of admissions in neurology department. Secondly, the clarity of the model structure is high, and the non-stationary properties of the original series of admissions are eliminated after trend difference (d = 1) and periodic difference (D = 1), indicating that the seasonal time series may need to be extended by SARIMA to clarify the interpretability of seasonal parameters. The lag order (p, q,P, Q) matches the periodicity of the data (such as the ACF/PACF diagram), indicating the interpretability of the ACF/PACF diagram for the lag order. Finally, the predicted results are consistent with the current knowledge and cognition in this field. Social pressures are increasing, which is leading to an overall rise in neurological disorders. Chinese folk customs believe that the Spring Festival should be reunited with the family, so for some less urgent neurological diseases, such as dizziness, headache, etc., Chinese people are more inclined to go to the hospital before the Spring Festival or after the Spring Festival, rather than go to the hospital during the Spring Festival.

As mentioned by Chai et al.³⁷, a combination of indicators is usually required to evaluate model performance, including but not limited to RMSE and MAE. Therefore, we comprehensively considered the principle that the model with the largest R²_adj and R² and the smallest MAPE, MAE and standardized BIC is the optimal model. The Winters addition model was selected among three exponential smoothing prediction models, and the SARIMA (0,1,1) (0,1,1) ₁₂ model was selected among several SARIMA models. Finally, we used the Winters addition model and the SARIMA (0,1,1) (0,1,1) ₁₂ model to predict admissions to neurology departments in October-December 2024, respectively, and found that the SARIMA (0,1,1) (0,1,1) ₁₂ model had a better predictive effect. That is, the MRE between the predicted value and the actual value is small, and the predicted value is closer to the actual value. The results of many reports on exponential smoothing and SARIMA model are similar to the results of this study^26,38,39. Swain et al.³⁸ used ARIMA and exponential smoothing method to predict the death toll of road accidents in India, and the results showed that ARIMA model had lower AIC and BIC values, which was superior to exponential smoothing method. Yang et al.²⁶ also found that compared with the seasonal exponential smoothing method, SARIMA model has a better prediction effect on the total number of hospital admissions. Mohammed et al.³⁹ adopted a variety of prediction models and found that in the seasonal autoregressive Comprehensive Moving average method (SARIMA), SARIMA exponential smoothing method (ETS), SARIMA neural network autoregressive method and SARIMA Adaptive Neural Fuzzy Reasoning System (SARIMA- anfis) model, SARIMA Adaptive Neural Fuzzy Inference System (SARMI-ANFIS) model has the best prediction effect.

In this study, the monthly data from January 2019 to September 2023 were used to establish a forecasting model. The annual seasonal cycle (12 months) was used, which actually included 4.75 complete seasonal cycles. We considered that in a rapidly changing medical environment, the use of recent 5-year data could better reflect the current diagnosis and treatment mode than the long-period data, avoiding the introduction of bias due to outdated historical data. In addition, a number of studies have reported 4–5 complete seasonal cycles to capture seasonal patterns to improve model stability and parameter estimation accuracy^29,30,31,32. We believe that the number of cycles in this study (4.75) has met the needs of basic modeling. Swedo et al.²⁹ used SARIMA model, generalized linear model, support vector machine and other models to predict emergency department visits and injury mortality of firearm injuries. This study selected data from 2014 to 2019, among which data from 2014 to 2017 were used for model training (four seasonal cycle data). The 2018 data is used for model validation, and the 2019 data is used as a test set to better evaluate the performance of the final model. Xian et al.³⁰ used the monthly incidence data of foodborne diseases of Nanan District Center for Disease Control and Prevention of Chongqing from June 2017 to June 2021 (four seasonal cycle data) to build three training models: SARIMA model, Holt-Winters model and exponential smoothing (ETS) model. And using July 2021 to April 2022 for prediction and validation, MSE, MAE, RMSE were used to determine the optimal model for SARIMA as SARIMA(1,0,0)(1,1,0)₁₂. To evaluate the performance of SARIMA model and Holt-Winters model in predicting the number of cases of Sjogren’s syndrome, Wang et al.³¹ took the monthly number of cases in the Department of Immunology and Rheumatology of Nanjing Zhongda Hospital from January 2015 to December 2019 as the training set (data from four seasonal cycles). Monthly incidence rates from January to December 2020 were used as a test set to test the performance of the model. Kalizhanova et al.³² used SARIMA and SIR Models to model the dynamics of tuberculosis transmission in Kazakhstan. The monthly TB notification rate from 2014 to 2018 was used as the data set for modeling (5 seasonal cycle data), and the 2019 data was used for forecasting. SARIMA model 1(2,1,0)(1,0,0)₁₂ and SARIMA model 2(2,1,0)(2,0,0)₁₂ were determined as the most suitable prediction models. Of course, we consider extending the data period beyond 8 years in future studies to better capture longer-term trends.

To verify the predictive effect of the model, we predicted the number of admissions in the department of neurology from October to December 2024. The results showed that compared with the Winters addition model, the average relative error between the actual and predicted number of admissions in the SARIMA (0,1,1) (0,1,1) ₁₂ model was smaller. This indicates that the model has a good predictive effect on the number of admissions in the department of Neurology, which is consistent with the results of Swain et al.³⁸ and Yang et al.²⁶. Both exponential smoothing model and SARIMA product season model are based on historical data and are widely used in forecasting time series data^{9,19,26,27,28,38,39}. However, compared with the two models, the exponential smoothing model cannot correct for the possible influencing factors, because the method forecasts the future trend through the weighted average of the historical time series data. SARIMA model can not only make full use of time series data change information, but also can incorporate possible influencing factors into the model for consideration. In our study, SARIMA model included possible influencing factors such as age of patients, average length of stay, Spring Festival, etc. into the model as control variables to eliminate their influence, which may also be one of the reasons for increasing the predictive efficiency of the model.

The results of this study suggest that age is a possible influencing factor for the change of admissions in neurology department. We included the data of inpatients in the Department of Neurology from 2019 to 2023. Common diseases include cerebral infarction, vertigo, arterial stenosis, transient ischemic attack, headache, epilepsy, Parkinson’s syndrome, etc. (Supplementary Table 1). Multiple studies have shown that common diseases in the department of neurology are age-related. Memon et al.⁴⁰ reported the correlation between old age and cerebral infarction. Wang et al.⁴¹ also found that age was an independent risk factor for the prognosis of cerebral infarction. A large epidemiological study in China has shown that cerebral infarction has a high incidence in people aged 40–70⁴². It may be because with the increase of age, the metabolic efficiency of the body decreases, and the elderly people have a higher risk of diabetes, hypertension, homocysteine and other metabolic diseases, and some studies have reported that diabetes, homocysteine and intracranial artery stenosis are related to cerebral infarction⁴⁰. Vertigo describes a symptom of abnormal movement of the environment or the patient’s own body, which affects the patient’s quality of life and daily activities, and increases healthcare utilization⁴³. Many studies have reported the relationship between age and vertigo^44,45. Studies based on large-scale populations have found that vestibular vertigo accounts for about a quarter of the main complaints of vertigo, and its prevalence increases with age, affecting the health of aging⁴⁴. Wassermann et al.⁴⁵ also found that the incidence of vertigo increased with age, resulting in significant limitations and disability in patients’ daily life. For arterial stenosis, a large study in South Korea found that the incidence of cerebral artery stenosis (CASTN) and the number of people treated were on the rise in recent years⁴⁶. Iranmanesh et al.⁴⁷ showed a significant relationship between age and cerebrovascular stenosis, and the elderly had a higher prevalence of cerebrovascular stenosis. Gutierrez et al.⁴⁸ also found that old age is a risk factor for intracranial atherosclerotic stenosis (ICAS). A transient ischemic attack (TIA) is defined as an ischemic attack in which the patient has neurological deficits but no acute infarction on imaging⁴⁹. Lip et al.⁵⁰ found that age is closely related to transient ischemic attack events. Sadighi et al.⁵¹ also found that age is a factor significantly related to the outcome of transient ischemic attack (TIA). For headache, a study of a general random population sample in Germany found that the incidence of headache decreased with age⁵². According to Rothermund and Brandstadter⁵³ ‘s theory on coping with deficits and losses in later life, we believe that it may be because psychological adaptation processes play an important role in the subjective perception of the impact of headache. Based on the longer duration of headache disorders, older headache patients may be better able to cope with their condition than younger people. More effective in generating emotional relief cognition. Epilepsy can be diagnosed in the presence of an unexplained seizure or epileptic syndrome. Epileptic syndromes represent a specific set of epilepsy types and EEG and imaging features, often with age-related features⁵⁴. A study based on global disease, injury, and risk factor data also found that the prevalence of active epilepsy increases with age, peaking at 5–9 years of age and over 80 years of age⁵⁵. A variety of laboratory rat models, including strains considered to be models of epilepsy deletion (such as Strasbourg gene deletion epileptic rats, GAERS), have explored why age is associated with epilepsy in old age, and the prevalence of spiking discharges increases with age, exploring this finding that may be related to epilepsy in old age. And this characteristic of naive old rats is crucial for the correct interpretation of experimental results⁵⁵. Parkinson’s disease, a progressive neurodegenerative disease characterized by tremors and motor retardation, is a common neurological disease. It is well known that old age is an independent risk factor, and aging is increasing the disease burden including Parkinson’s disease^56,57. The global burden of Parkinson’s disease more than doubled between 1990 and 2016 due to an increase in the number of older adults, and other underlying factors such as longer disease duration and environmental factors may be responsible for the increased burden of Parkinson’s disease in the future⁴⁷. Chen et al.⁵⁸ found that age has different degrees of influence on the brain structural network and cognitive function of Parkinson’s disease patients. Compared with the low age group, Parkinson’s disease patients in the high age group show the interruption of white matter network topology and the impairment of white matter fiber integrity. In recent years, China’s aging has become more and more serious, and a number of studies have reported that old age is an influential factor for common diseases in the department of neurology, which may be the reason for the overall upward growth in the number of admissions to the department of neurology. We speculate that the long-term demand for neurology will continue to rise in the future, and the pressure on medical resources will be further upgraded. In this regard, we suggest the integration of hospital specialties and multi-disciplines. Specialties are to set up a sub-specialty of geriatric psychiatry and formulate treatment norms according to the characteristics of elderly patients (such as cognitive impairment and drug metabolism differences). Multidisciplinary integration is the normalization of a multidisciplinary team (MDT), combining neurosurgery, geriatrics, rehabilitation, psychology and other departments to optimize complex case management (such as early intervention of post-stroke depression). Renovate the aging infrastructure and build age-friendly wards.

In addition, some studies have suggested that the number of beds may be a possible confounding factor for changes in admissions⁹. However, the number of beds in the neurology department did not change in the Tier 3 hospitals included in this study, so we did not include the number of beds as a possible influencing factor in the model. As for socioeconomic status, the data in this study came from the hospital HIS system, which was limited by the variable dimension of single-center data (including only age, spring factor, diagnosis category, etc.), and this variable could not be obtained, which was data limitations. As for comorbidities, many studies have reported the common comorbidities of nervous system diseases, such as hypertension, diabetes, atrial fibrillation, depression, hyperlipidemia, metabolic syndrome, Osteoporosis and cognitive disorders^{59,60,61,62,63,64,65,66,67,68,69,70,71}. Global data have confirmed that hypertension and high fasting glucose are important risk factors for stroke⁵⁹. Hypertension is a global public health problem and one of the leading risk factors for premature death and disability from stroke⁶¹. A prospective nested case-control study conducted by Yang et al.⁶¹ found that type 2 diabetes was significantly associated with increased risk of cerebral infarction and cerebral artery occlusion, and diabetes was a risk factor for stroke⁶². Simon et al.⁶³ found that post-stroke depression (PSD) is a clinically related complication of ischemic and hemorrhagic cerebral infarction, accounting for about 30% of all stroke survivors. Seiffge et al.⁶⁴ reported that atrial fibrillation is one of the most common arrhythmias and the main cause of ischemic stroke. For arterial stenosis, Miura et al.⁶⁵ showed that hypertriglyceridemia would increase the risk of the progression of carotid artery stenosis, and lipid-lowering drugs played a potential role in carotid artery stenosis. Gutierrez et al.⁶⁶ found that intracranial atherosclerotic stenosis (ICAS) is one of the most common causes of stroke worldwide. Compared with other causes of stroke, ICAS is one of the greatest risks of stroke recurrence. Shu et al.⁶⁷ found that the risk of ICAS with metabolic syndrome (MetS) severity score was 1.75(OR = 1.75, 95% CI 1.39–2.21). In addition, people with epilepsy have an increased risk of vertebral and non-vertebral fractures, which may be due to osteoporosis, increased bone fragility, and increased risk of falls, and it has been established that anti-epileptic drugs (AEDs) can also have specific effects on bone metabolism⁶⁸. For Parkinson’s syndrome, neuropsychiatric symptoms (nps) such as cognitive impairment are common in Parkinson’s disease (PD)⁶⁹. Cognitive impairment in patients with PD varies in its clinical characteristics and rate of progression, and is now recognized to occur throughout the disease, including from early, de novo to more advanced stages⁷⁰. van et al.⁷¹ reported that PD is an independent risk factor for osteoporosis. Therefore, the ICD-10 code in the electronic medical record was extracted to analyze the frequency of complications from January 2019 to September 2023 (the diagnosis code of ICD-10 disease includes hypertension, diabetes, atrial fibrillation, depression, hyperlipidemia, metabolic syndrome, Osteoporosis and cognitive disorders). The results showed no correlation between the rate of comorbidity and the number of admissions to the department of Neurology (P > 0.1) (see Table 7). Therefore, we did not include the rate of comorbidity as a control factor in the model. As for hospital policies, the policy of the neurology department included in the hospital in this study did not change much, nor did the number of beds and other configurations, so we did not include the variable of hospital policies.

For the change of healthcare policy, we expect to use interrupted time series (ITS) to analyze, ITS analysis can examine how the intervention affects the sequence, control the effect of the regression trend before the intervention on the sequence. The effectiveness of the intervention measures was evaluated by comparing and testing the immediate level change of the outcome variables before and after the intervention and the slope change of the regression line before and after the intervention⁷². In addition, studies have reported that ITS needs to collect 40 to 50 data points, or at least 20 points before intervention and 20 points after intervention⁷³. However, our study only had 12 data before the implementation of the medical insurance DRGs payment policy (The Medicare DRGs payment policy was implemented in January 2020), which failed to meet the requirements of analysis. In the future, we consider including more variables in the analysis. However, although variables such as medical insurance DRGs payment policy, public health campaign and epidemiological trend are not included, they are only possible confounding factors. We believe that changes in medical insurance policy only change the payment method, and will not affect residents’ medical treatment behavior, and public health campaign and epidemiological trend will not affect sick residents to go to hospital too much. In addition, among the variables included in our study, the results were statistically significant, and the number of admissions in the department of neurology showed a periodic and regular change, which has practical guiding significance for the reasonable allocation of medical resources, the formulation of reasonable policies by decision makers, the provision of better services for residents, and even the improvement of patient satisfaction.

Furthermore, it is important to emphasize that this study only focused on the neurology department of one hospital in Zhejiang Province, and there may be some limitations to the generalization of these findings to other departments or hospitals, especially those with different patient demographics, healthcare systems, or seasonal patterns. Therefore, to explore the changes in the number of admissions in other departments or hospitals, parameters should be adjusted according to the localization characteristics.

Nevertheless, potential obstacles to the successful application of predictive models in hospital Settings still need to be addressed. First, the compatibility of real-time data interfaces needs to be addressed, such as the standardization of the data format of the Hospital Information system [HIS] system; Secondly, we have established a multi-center hospital data integration system, which integrates the medical record management system, nursing system, surgery system and anesthesia system into a unified data index platform (see Supplementary Fig. 1). In the future, we will continue to use this data integration system to promote the successful application of prediction model in hospital environment. In addition, we will first adopt lightweight deployment solutions in the future, such as developing visualization tools with low computing requirements based on RShiny, and then gradually promote them. Finally, personnel training is crucial. In our study, the medical department took the lead in the application plan of the prediction model in the hospital environment, and the hospital Information Department and teaching Department jointly designed the training module to train the doctors, nurses and medical technicians in the hospital to operate the data integration system proficiently and update the data regularly.

In addition, the use of predictive models in health care may have some ethical implications, potential biases or predictions in the data may affect the allocation of patient care and resources, there are issues of too much or too little resource allocation due to data biases, affecting good patient care, and there may be unreasonable formulation of hospital policies based on data biases. To solve this problem, first of all, the prediction model is still in the testing stage, and is more used for scientific research and theoretical experiments. We will continuously optimize the prediction model, improve the prediction efficiency and accuracy of the model, and make the prediction accuracy stable to more than 95%. Then start the trial operation stage in the hospital, in this stage, the medical staff only based on the forecast data as a reference, reasonable allocation of medical resources, at the same time, more than 10% of the flexible medical resources, to ensure that patients get more adequate and better care.

In short, according to the changes of patients admitted to the department of Neurology in a top-three hospital in Zhejiang Province, we suggest that the decision makers of the hospital conduct the overall medical resource allocation from the following aspects. On the one hand, the hospital should improve the utilization rate of idle resources and reduce hospital costs. For example, it is suggested that the health department should allocate flexible beds in the neurology department and share the vacant beds with other departments in the hospital in January and February. By using idle resources, the waiting time of patients can be reduced, the time cost of patients can be reduced and patient satisfaction can be improved. On the other hand, it is suggested that hospitals should integrate specialties and disciplines. Specialties are to set up a sub-specialty of geriatric psychiatry, and formulate diagnosis and treatment norms according to the characteristics of elderly patients (such as cognitive impairment and drug metabolism differences). Multidisciplinary integration is the normalization of multidisciplinary teams (MDT), which combine neurosurgery, geriatrics, rehabilitation, psychology and other departments to optimize complex case management. In addition, it is recommended to embed the admission prediction module in the hospital system to quantify resource allocation adjustments, predict the changes in the number of admissions in the neurology department in real time, and deal with the issues of real-time data integration, computing resources or staff training in the implementation process, which are crucial for the successful application of the prediction model in the hospital environment.

Conclusion

We concluded that both the SES model and the SARIMA model could be used to predict the number of admissions in the neurology department, and the SARIMA (0,1,1) (0,1,1) ₁₂ model had a better predictive effect on the number of admissions than the Winters addition model. We suggest that the hospital management departments rationally arrange medical resources according to the annual cyclical changes.