Comparison of ARIMA model and LSTM model in predicting disease burden of occupational pneumoconiosis in Tianjin, China

doi:10.21203/rs.3.rs-1410451/v1

PDF

Research Article

Comparison of ARIMA model and LSTM model in predicting disease burden of occupational pneumoconiosis in Tianjin, China

https://doi.org/10.21203/rs.3.rs-1410451/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

To evaluate the epidemic trend of the disease burden of occupational pneumoconiosis in Tianjin and explore appropriate mathematical models to predict the disease burden of pneumoconiosis.

Methods

Disability adjusted life year (DALY) was used to evaluate the disease burden of occupational pneumoconiosis in Tianjin from 1990 to 2019. ARIMA model and LSTM model were used to fit and predict the DALY of pneumoconiosis. Five performance evaluation indexes including Mean Square Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) were selected to compare the prediction effects of the two models.

Restults:

DALY caused by pneumoconiosis in Tianjin showed a fluctuating upward trend from 1990 to 2019, with three peaks in 1995, 2005 and 2015. ARIMA (0, 2, 1) model was determined as the best fitting ARIMA model. The LSTM model with 6 neurons and Adaptive Moment Estimation (Adam) was selected as the best fitting LSTM models. The MSE, RMSE, MAE and MAPE values of LSTM model were lower than those of ARIMA model, indicating that LSTM model had a good prediction effect on pneumoconiosis DALY. The LSTM model was used to predict that the DALY of pneumoconiosis in Tianjin showed a downward trend from 2020 to 2023, which was 2873.82, 2550.19, 2291.02 and 2069.68 person-years, respectively.

Conclusion

LSTM model has good performance in predicting the disease burden of pneumoconiosis and has great application prospect in the field of occupational diseases.

Pneumoconiosis

DALY

ARIMA model

LSTM model

Pneumoconiosis is a group of heteroeous occupational interstitial lung diseases related to the corresponding reactions of inhaled mineral dust and lung tissue, which eventually leads to irreversible lung injury[1]. Due to the lack of prevention of workplace dust, failure of early diagnosis of diseases, and limited effective treatment of diseases, Pneumoconiosis is still a serious global public health problem. According to the Global Burden of Disease(GBD) Study 2017[2], the number of pneumoconiosis cases worldwide increased by 66.0% from 36,186 in 1990 to 60,055 in 2017 and approximately 125,000 cases of pneumoconiosis die every year[3]. China is the country with the largest working population in the world and the most seriously affected by pneumoconiosis. In 2018, it was about 776 million working population, and most workers spend half their lives working. According to the estimation of the National Health Commission of China, the total number of occupational-based cases reported by 2018 was 97500, and 90% of the occupational diseases were identified as pneumoconiosis[4]. Tianjin is an industrial city dominated by processing and manufacturing industries in northern China. In recent years, the number of cases of pneumoconiosis in Tianjin has been increasing, showing a trend of wide industrial distribution and younger age. Although China has taken a variety of measures to prevent and control pneumoconiosis in the past few decades, compared with the United States and Britain, China 's occupational health field is still in its infancy, and the situation of pneumoconiosis prevention and control is still grim. Pneumoconiosis causes huge disease burden and economic losses to Chinese workers, families and society every year.

Disease burden assessment is an important public health tool to guide risk reduction and prevent diseases caused by workplace exposure. Disability adjusted life year (DALY) was developed by WHO and the World Bank to quantify human disease burdens and injuries in the Global Burden of Disease Study[5]. DALY combines the estimation of disability survival time and the time lost due to premature death, and is adjusted through a set of social preference values[6]. For different age groups and time periods, DALY can be given different age weights and discount rates. Therefore, this provides an objective and quantitative description of the gap between ideal health status and actual population health status[7]. Due to these irreplaceable advantages, DALY method has been applied in many fields, such as cancer[8], cardiovascular diseases[9], and the impact of environmental pollution on health[10]. However, it is relatively less applied in the field of occupational diseases.

Also known as historical extension forecasting method, time-series forecasting method is an extrapolation and forecasting method to reflect the development trend of things through time-series[11]. Accurately predicting the trend of pneumoconiosis burden can realize the monitoring and early warning of pneumoconiosis. Common traditional time-series prediction methods include autoregressive integrated moving average (ARIMA) model and Holt-Winters exponential smoothing method et al, among which ARIMA model is the most classical and popular model[12, 13]. ARIMA model involves the invariance of trend change, random disturbance, periodic change and other related random variables in the process of time-series analysis. Due to the advantages of simple structure, strong applicability and ability to interpret data sets, ARIMA model has been successfully applied in the past medical and health fields[14].

In recent years, deep learning technology has developed rapidly and is widely used to extract information from various data. Among them, Artificial neural network (ANN) is widely used because it can overcome the limitation of linear model.[15]. In terms of time-series model prediction, recurrent neural networks (RNN) model dominates and has higher prediction accuracy than traditional artificial neural network[16, 17]. However, when the sequence length is too large, the training time of RNN is significantly increased and it is prone to gradient disappearance and gradient explosion[18]. Based on the above problems, a novel recursive network structure called Long Short-Term Memory Neural Network (LSTM) was proposed[19]. It combines the appropriate gradient-based learning algorithm, improves the hidden layer of RNN and extends the storage function of the network, so that the model can obtain more persistent information and reduce data transmission speed[20, 21]. It can learn to span 1000 steps without losing short latency even in noisy, incompressible input sequences[22]. In recent years, LSTM model has been more and more applied in many fields such as traffic flow prediction, speech recognition and disease prediction[23, 24]. As far as we know, no studies using LSTM model to predict the disease burden of pneumoconiosis.

In this study, ARIMA model and LSTM model were used to fit and predict the time-series of pneumoconiosis burden in Tianjin, China. In addition, by comparing the fitting effect and prediction accuracy of the two models, a more suitable prediction model is sought to predict the disease burden level of pneumoconiosis in Tianjin

Data source

The case data of pneumoconiosis in this study were collected from the follow-up survey of occupational pneumoconiosis patients in China ' s National Programme of Action for the Prevention and Treatment of Pneumoconiosis. The basic information of pneumoconiosis patients in 2005 and before was obtained by the epidemiological survey data of pneumoconiosis, and the data of occupational pneumoconiosis cases reported from 2006 to 2019 was obtained by the occupational disease reporting system.

The basic information of pneumoconiosis patients such as gender, age, survival, region, industry classification, dust exposure time, pneumoconiosis type, stage, diagnosis date, death date and other information were collected.

DALY Calculation

DALY can be defined as the total loss of healthy life years from onset to death[25], which consists of Years of Life Loss (YLLs) due to premature mortality and Years Live with Disability (YLDs) due to disability[7]. Several social preference values should be considered in the calculation of DALY, such as the disability weight between 0 and 1, which reflects the severity of health hazards caused by different diseases. The age weight is used to distinguish the relative life value of different age groups[26] and the time discount rate to distinguish the relative value of health life loss occurs in different periods. Considering these factors, the complete formula for calculating DALY can be expressed as integral form[6]

$${\int }_{x=a}^{x=a+L}DCx{e}^{-\beta x}{e}^{-\gamma (x-a)}dx$$

Each parameter is determined by WHO Global Burden of Disease (GBD). Where D is disability weight (1 is premature death), γ is discount rate (0.03), C is age correction weight value (0.16243); β is the age weighting function (0.04), a is the starting age, L is the disability time or premature death time[27].

The definite integral formulas of YLL and YLD are obtained by function integral:

YLLs=$\frac{KC{e}^{r\alpha }}{{(r+\beta )}^{2}}\left\{{e}^{-(r+\beta )(L+a)}\left[-\left(\text{r}+\beta \right)\left(L+\alpha \right)-1\right]-{e}^{-\left(r+\beta \right)a}[-\left(r+\beta \right)\alpha -1]\right\}+\frac{1-K}{r}(1-{e}^{-rL})$ (2)

YLDs=$D\frac{KC{e}^{r\alpha }}{{(r+\beta )}^{2}}\left\{{e}^{-(r+\beta )(L+a)}\left[-\left(\text{r}+\beta \right)\left(L+\alpha \right)-1\right]-{e}^{-\left(r+\beta \right)a}[-\left(r+\beta \right)\alpha -1]\right\}+\frac{1-K}{r}(1-{e}^{-rL})$ (3)

ARIMA model

ARIMA model has two parts: autoregressive (AR) and moving average (MA). In general, the model is expressed as ARIMA (p, d, q), p means the order of auto-regression, d means the order of difference and q means the order of moving average[28]. ARIMA needs to transform the non-stationary time-series into a stationary time-series, and then a model is established by regression of the lag value of the dependent variable and the present value and lag value of the random error term. The basic idea is to regard the data formed by the predicted object over time as a random sequence, describe the autocorrelation in the sequence with the corresponding mathematical model, and predict the future value by using the potential relationship between the past value and the present value of the sequence. The three main steps of establishing ARIMA time-series model are as follows: (1) Data preprocessing, observing the time-series diagram, autocorrelation analysis diagram and using the Augmented Dickey-Fuller (ADF) unit-root test to estimate whether the time-series is stable. If the sequence is a non-stationary sequence, the corresponding difference is used to smooth the sequence, and white noise test is carried out to test whether the difference sequence is white noise sequence; (2) Model identification, order determination and model parameter estimation. Autocorrelation Function (ACF) graph and Partial Autocorrelation (PACF) graph are used to estimate parameters, and the optimal model types and parameters can be screened by combining Akaike information criterion (AIC) and Bayesian information criterion (BIC), usually with the lowest AIC or BIC values[29]; (3) The Q-Q plots are used to test whether the residuals of the model meet the independent normal distribution, and the white noise analysis of the residuals is used to diagnose and test the optimal model. Finally, the better fitting model is used to predict[30].

LSTM model

LSTM is a machine learning algorithm with recursive neural network structure[31]. According to the defined parameters and algorithms, LSTM neural network adds a gate structure to control the state of memory cells in each neuron, which is input gate, output gate and forget gate. As shown in Fig. 1, the first forgetting gate selectively forgets the input from the previous node, and the data flowing to the gate are the input ${x}_{t}$ at the current time and the output ${h}_{t-1}$ of the hidden layer at the previous time; the input gate still processes ${x}_{t}$ and ${h}_{t-1}$, selectively memorize the data; Finally, the output gate further processes the data to obtain the output of the hidden layer ${h}_{t}$. The basic formula is as follows:

Forgetting gate expression:${f}_{t}$=σ (${W}_{1}^{f}$·${x}_{t}$+${W}_{h}^{f}$·${h}_{t-1}$+${b}_{f}$); Input gate expression: ${i}_{t}$=σ (${W}_{1}^{i}$·${x}_{t}$+${W}_{h}^{i}$·${h}_{t-1}$+${b}_{i}$); The expression of the current input unit state ${\tilde{C}}_{t}$=tanh(Wc·${h}_{t-1}$+Wc·${x}_{t}$+${b}_{c}$); The current unit state is the last unit state multiplied by the element to the forgetting gate, plus the current input unit state multiplied by the element to the input gate: ${C}_{t}$=${f}_{t}{*C}_{t-1}$+${i}_{t}*{\tilde{C}}_{t}$; Output gate expression: ${o}_{t}$=σ (${W}_{1}^{0}$·${x}_{t}$+${W}_{h}^{0}$·${h}_{t-1}$+${b}_{0}$); Final output of LSTM model: ${\text{h}}_{t}$=${o}_{t}$*tanh (${C}_{t}$). Where ${i}_{t}$, ${f}_{t}$, ${o}_{t}$ and ${C}_{t}$ represent the corresponding three gate structures and cell states at time t, W is a weight matrix connecting two layers, b is a bias, and σ is the sigmoid or tanh function[32–34].

Model Comparison

Four performance indicators including mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) were used to compare and evaluate the fitting and prediction accuracy of the two models. MAE is the simplest measure of fitting and prediction accuracy that determines the average prediction error. MAPE is the mean value of unsigned percentage error, which can solve the problem of distinguishing large error from small error, but it may underestimate the rare error. Therefore, we can adjust the larger rare error by calculating the MSE and RMSE index. The calculation method of MSE is to square the error first, and then calculate the average error. The calculation of RMSE takes the square root on the basis of MSE calculation. The specific calculation formula is as follows[35]:

MAE=$\frac{1}{n}{\sum }_{i=1}^{n}$|${\widehat{y}}_{i}-{y}_{i}$| (4)

RMSE=$\sqrt{\frac{1}{n}{\sum }_{i=1}^{n}({\widehat{y}}_{i}-{y}_{i})2}$ (5)

MSE=$\frac{1}{n}{\sum }_{i=1}^{n}$(${\widehat{y}}_{i}-{y}_{i}$)² (6)

MAPE=$\frac{100\%}{n}{\sum }_{i=1}^{n}$|$\frac{{\widehat{y}}_{i}-{y}_{i}}{{y}_{i}}$| (7)

Where ${\widehat{y}}_{i}$ is the predicted value, ${y}_{i}$ is the actual value, and n is the number of predicted data. The smaller the above index values, the higher the prediction accuracy of the model.

Statistical Analysis

Excel 2019 software was used to establish a database, and the World Health Organization disease burden Excel template was used to calculate the DALY of pneumoconiosis. The data from 1990 to 2016 were used as the training set, and 2017–2019 were used as the testing set to establish the prediction model. Python 3.9.5 is used to establish ARIMA model and LSTM neural network model. ARIMA model is mainly realized by statsmodels library, and LSTM model is mainly constructed based on PyTorch framework library of Anaconda environment. In this study, the statistical significance level of all hypothesis tests was set to 0.05.

Trend of pneumoconiosis burden in Tianjin

As presented in the Fig. 2, DALY caused by pneumoconiosis in Tianjin showed a fluctuating upward trend from 557 person-years in 1990 to 3333 person-years in 2019. There were three peaks in 1995, 2005 and 2015, respectively 2025 person-years, 4207.451465 person-years and 5306 person-years. Among them, the increase rate from 2004 to 2005 was the largest, and it showed a significant downward trend after 2017.

Fitting models with ARIMA

Sequence stabilization

The original sequence diagram of the training set showed a fluctuating trend (Fig. 3a), The ADF unit-root test showed that t = 0.834, P = 0.806, which could not reject the original hypothesis. Therefore, the sequence could be determined to be non-stationary according to the above information, and differential processing is needed.

The sequence diagram of the original sequence tended to be stable after twice difference (Fig. 3b). The ADF unit-root test suggested that t = − 7.644, P < 0.0001, rejecting the original hypothesis and meeting the requirements of sequence stability. Therefore, the parameter d was set to 2

Model identification and screening

The ACF (Fig. 3c) and PACF (Fig. 3d) diagrams of time-series showed that the ACF does not drop rapidly to 0 after several orders of lag, and there is obvious trailing phenomenon. The PACF decayed rapidly after the first order, and fluctuated in a small range around the zero axis and basically falls within the confidence interval. According to the Bayesian information criterion, the model with the minimum BIC value is optimal. As shown in Fig. 4a, AR represents p, MA represents q. When p is 0 and q is 1, the minimum BIC value is 402.25, so the optimal model is ARIMA (0,2,1).

Model test and prediction

As shown in Fig. 4b, the quantile plot method (Quantile-Quantile Plot, Q-Q plot) was used to prove that the residual of the model conforms to the normal distribution. The D-W test suggested that the D-W value was 2.137 close to 2, indicating that there was no autocorrelation. The test results of residual white noise (Ljung-Box) showed that P = 0.528 > 0.05, indicating that the residual is a white noise sequence. The above tests showed that ARIMA (0, 2, 1) model is an effective model that meets the requirements.

Fitting models with LSTM

Data normalization processing

In order to improve the convergence speed and fitting accuracy of the model, the minimum and maximum standard 'MinMaxScaler ( )' was used to convert the original data to 0 ~ 1, and the standardized data was used to model. Finally, the results of the model output are restored.

Establishment of LSTM Model

The LSTM model established in this study has 2 input layers, 6 neurons in the hidden layer and the output layer as the predicted value. The sigmoid function was used as the activation function. The optimizer used adam, batch_size was set to 1, the output layer was set to linear function tanh for output. The number of iterations was set to 100. Mean_squared_error was used to calculate the loss function value of each step of training and the loss value decreased with the increase of training times (Fig. 5). In order to prevent the over-fitting of the training set, L2 regularization was adopted and Dropout function was added between the hidden layers. The model adjusted the value of look_back to find the optimal situation of the current network structure.

Prediction results

The comparison results between the predicted values and the actual values of the two models are shown in Table 1 and Fig. 6. In the figure, the actual DALY values of pneumoconiosis in Tianjin from 1990 to 2019 were shown by the blue line and the predicted values were represented by red line. It can be seen from the figure that the predicted values of the LSTM model are closer to the actual value.

Table 1

Comparison of predicted and actual values of ARIMA model and LSTM model
Year	Actual value	ARIMA Predictive value	LSTM Predictive value
1992	809.148	734.465	766.994
1993	644.751	881.906	693.364
1994	1575.736	794.800	1551.854
1995	2025.294	1850.985	2039.298
1996	1316.641	2231.616	1314.520
1997	575.491	1378.873	572.908
1998	593.512	579.121	599.236
1999	1190.540	631.955	1188.995
2000	1249.609	1260.348	1252.145
2001	1724.612	1337.749	1722.420
2002	1812.652	1829.985	1824.170
2003	1594.135	1904.609	1584.276
2004	2156.407	1691.463	2166.339
2005	4207.451	2336.321	4210.928
2006	4052.754	4438.173	4076.587
2007	4194.155	4268.935	3938.381
2008	3350.975	4378.319	3554.789
2009	3847.568	3515.299	3782.591
2010	3843.085	4016.255	3732.414
2011	3979.936	4006.726	3718.295
2012	3864.620	4136.533	3790.705
2013	3604.896	4005.951	3785.027
2014	4935.472	3762.637	4675.913
2015	5305.733	5121.702	5063.327
2016	4869.672	5483.524	4952.507
2017	5186.590	4715.124	5166.283
2018	4165.548	4462.291	4168.240
2019	3333.095	4243.773	3668.210

Model comparison

As shown in Table 2, the MAPE, MAE, RMSE and MSE of the training set in LSTM model were 3.2014, 0.0158, 0.1257 and 0.0006, respectively, which are significantly lower than those of ARIMA model. The four evaluation indexes of the testing set in LSTM model were 4.2187, 0.0243, 0.1559 and 0.0002, which are also significantly lower than those of ARIMA model.

Table 2

Comparison of ARIMA and LSTM models in predicting DALY of pneumoconiosis
Performance index	Training set		Testing set
Performance index	ARIMA	LSTM	ARIMA	LSTM
MAPE	40.6247	3.2014	17.1573	4.2187
MAE	0.0919	0.0158	0.1140	0.0243
RMSE	0.3032	0.1257	0.3376	0.1559
MSE	0.0163	0.0006	0.0158	0.0002

Prediction of DALY for future pneumoconiosis in Tianjin

Combined with the calculation results of evaluation indexes of training set and test set, the performance and accuracy of LSTM model in predicting DALY of pneumoconiosis were better than those of ARIMA model. Therefore, LSTM model was used to fit the disease burden from 1990 to 2019, and the DALY of pneumoconiosis in Tianjin from 2020 to 2023 was predicted to be 2873.82, 2550.19, 2291.02, 2069.68 person-years, showing a downward trend.

Pneumoconiosis is the most serious occupational disease that endangers the health of workers in Tianjin, causing a huge disease burden every year. In this study, DALY was used to quantitatively analyze the disease burden of pneumoconiosis patients in Tianjin from 1990 to 2019. The application of mathematical models to accurately predict the development trend of pneumoconiosis burden is helpful for health decision makers to grasp the status of pneumoconiosis burden, so as to make objective judgments as soon as possible and take effective prevention and control measures. As far as we know, this study first used ARIMA model and LSTM model to establish a prediction model of pneumoconiosis burden in Tianjin.

ARIMA model is a classical time-series model developed on the basis of linear regression model, which combines the advantages of autoregressive model and average moving model[36]. It can reveal the dynamic law of data and unify the comprehensive effect of influencing factors into the time variable, which can not only avoid the influence of factors related to disease burden or the difficulty of obtaining data, but also overcome the random interference problem. However, it has strict requirements on data and requires data to meet stationary sequences or stationary sequences after differential conversion. Moreover, the model identification and calculation are relatively complex, and there are problems such as weak nonlinear mapping performance and difficult to fit irregular sequences[28, 37]. In this study, the DALY of pneumoconiosis was non-periodic and seasonal data, and the fluctuation range of data was large. It was necessary to use the quadratic difference to meet the requirements of the stationary sequence, but the difference data generated the corresponding information loss. The results show that the fitting effect of ARIMA model can basically reflect the fluctuation trend of DALY of pneumoconiosis in Tianjin, but the MAPE of the testing set and the training set of ARIMA model were higher than 10%. It is generally believed that MAPE < 10% indicates that the model has strong prediction efficiency[38]. Therefore, the effect of ARIMA model in predicting the disease burden of pneumoconiosis is general.

LSTM is an advanced recurrent neural network that aims to mine information from data itself, learn time patterns and capture nonlinear dependencies[39]. In the model, each neuron calls information circularly and transmits it to the next neuron. At the same time, the weights are adjusted by adding or subtracting information to avoid the problems caused by long-term sequences and store useful memory in a longer time. Therefore, it can produce better prediction results when the number of data sets is large, and it is more suitable for data with large fluctuations[31, 33]. The results of this study suggested that all the evaluation indexes of LSTM model were lower than those of ARIMA model, and MAPE was lower than 10%, indicating that LSTM neural network model has high application value for the prediction of pneumoconiosis disease burden.

The LSTM model was used to predict that the disease burden of pneumoconiosis in Tianjin will decrease in the next four years. To a certain extent, this prediction result shows that with the joint efforts of occupational disease prevention and control institutions and society at all levels, the prevention and control of pneumoconiosis in Tianjin has achieved remarkable results, and the disease burden caused by pneumoconiosis will be improved. However, pneumoconiosis is still the main occupational disease in Tianjin, and the disease burden of pneumoconiosis is still at a high level throughout the country. The next 10 years is an important strategic opportunity to promote the construction of healthy China, and strengthening the study of disease burden is an important measure. In the future, with the help of more mathematical prediction models, we will improve the monitoring and early warning of pneumoconiosis disease burden and effectively improve the level of occupational disease prevention and control.

The LSTM neural network model has better overall performance than the traditional ARIMA model in predicting the disease burden of pneumoconiosis, and it will have great application prospects in the field of occupational diseases in the future. However, there are still some shortcomings in this study. This study only establishes the model from the perspective of single variable time-series without considering other relevant factors, which may lead to large deviations with external changes. However, with the development of prediction theory and technology, more prediction methods and appropriate models will be applied to the prediction and early warning of occupational diseases, so as to achieve a more scientific and accurate prediction of future trends.

DALY

Disability adjusted life year

ARIMA

autoregressive integrated moving average

LSTM

Long Short-Term Memory Neural Network

MSE

Mean Square Error

RMSE

Root Mean Squared Error

MAE

Mean Absolute Error

MAPE

Mean Absolute Percentage Error.

Acknowledgments

The authors would like to sincerely thank Tianjin Center for Disease Control and Prevention for the support of this study and the Establishment and maintenance of the Public Health Information System database of the Chinese Center for Disease Control and Prevention. The authors would also like to thank anonymous peer reviewers for carefully revising our manuscript and for his or her useful comments

Authors’ contributions

LHR participated in design, data analysis and drafting the manuscript. GY participated in follow-up surveys and data collection. WX explained the data and critically reviewed the intellectual content of the article. ZQ provided technical support and helped revise the manuscript. The final manuscript has been read and approved by all the authors.

Funding

This work was supported by grants from Tianjin Natural Science Foundation of China (No.20JCYBJC00270) and Medical Key Discipline (Specialized Discipline) Construction Project of Tianjin (2021).

Availability of data and materials

The datasets generated and/or analysed during the current study are not publicly available due to the personal privacy contained in the data but are available from the corresponding author on reasonable request.

Ethics approval and consent to participate

Ethical permission for the study is not essential since it is the monitoring data of pneumoconiosis in government functions. The investigators have gotten the survey data from the follow-up questionnaire of occupational pneumoconiosis program and then the investigators of this study have preserved the privacy of the data. Informed consent was obtained from all participants. All methods were performed in accordance with the Follow-up Investigation Program of Occupational Pneumoconiosis.

Consent for publication

Not applicable

Competing interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Author details

¹School of Public Health, Tianjin Medical University, Tianjin 300070, China.²Tianjin Center for Disease Control and Prevention, Tianjin, 300011, P. R. China.

Qi XM, Luo Y, Song MY, Liu Y, Shu T, Liu Y, Pang JL, Wang J, Wang C: Pneumoconiosis: current status and future prospects. Chin Med J (Engl) 2021, 134(8):898–907.
Rudd KE, Johnson SC, Agesa KM, Shackelford KA, Tsoi D, Kievlan DR, Colombara DV, Ikuta KS, Kissoon N, Finfer S et al: Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study. LANCET 2020, 395(10219):200–211.
Lozano R, Naghavi M, Foreman K, Lim S, Shibuya K, Aboyans V, Abraham J, Adair T, Aggarwal R, Ahn SY et al: Global and regional mortality from 235 causes of death for 20 age groups in 1990 and 2010: a systematic analysis for the Global Burden of Disease Study 2010. LANCET 2012, 380(9859):2095–2128.
The L: Improving occupational health in China. LANCET 2019, 394(10197):443.
Murray CJ: Quantifying the burden of disease: the technical basis for disability-adjusted life years. Bull World Health Organ 1994, 72(3):429–445.
Gao T, Wang XC, Chen R, Ngo HH, Guo W: Disability adjusted life year (DALY): a useful tool for quantitative assessment of environmental pollution. SCI TOTAL ENVIRON 2015, 511:268–287.
Murray CJ, Lopez AD: The utility of DALYs for public health policy and research: a reply. Bull World Health Organ 1997, 75(4):377–381.
Di Pardo BJ, Bronson NW, Diggs BS, Thomas CJ, Hunter JG, Dolan JP: The Global Burden of Esophageal Cancer: A Disability-Adjusted Life-Year Approach. WORLD J SURG 2016, 40(2):395–401.
Wafa HA, Wolfe C, Emmett E, Roth GA, Johnson CO, Wang Y: Burden of Stroke in Europe: Thirty-Year Projections of Incidence, Prevalence, Deaths, and Disability-Adjusted Life Years. STROKE 2020, 51(8):2418–2427.
Cohen AJ, Brauer M, Burnett R, Anderson HR, Frostad J, Estep K, Balakrishnan K, Brunekreef B, Dandona L, Dandona R et al: Estimates and 25-year trends of the global burden of disease attributable to ambient air pollution: an analysis of data from the Global Burden of Diseases Study 2015. LANCET 2017, 389(10082):1907–1918.
Li Z, Li Y: A comparative study on the prediction of the BP artificial neural network model and the ARIMA model in the incidence of AIDS. BMC Med Inform Decis Mak 2020, 20(1):143.
Wang Y, Xu C, Yao S, Zhao Y: Forecasting the epidemiological trends of COVID-19 prevalence and mortality using the advanced alpha-Sutte Indicator. EPIDEMIOL INFECT 2020, 148:e236.
Adeyinka DA, Muhajarine N: Time series prediction of under-five mortality rates for Nigeria: comparative analysis of artificial neural networks, Holt-Winters exponential smoothing and autoregressive integrated moving average models. BMC MED RES METHODOL 2020, 20(1):292.
Ceylan Z: Estimation of COVID-19 prevalence in Italy, Spain, and France. SCI TOTAL ENVIRON 2020, 729:138817.
Zaini N, Ean LW, Ahmed AN, Malek MA: A systematic literature review of deep learning neural network for time series air quality forecasting. Environ Sci Pollut Res Int 2022, 29(4):4958–4990.
Escudero P, Alcocer W, Paredes J: Recurrent Neural Networks and ARIMA Models for Euro/Dollar Exchange Rate Forecasting. APPL SCI-BASEL 2021, 11(12).
Wu W, An S, Guan P, Huang D, Zhou B: Time series analysis of human brucellosis in mainland China by using Elman and Jordan recurrent neural networks. BMC INFECT DIS 2019, 19.
Gu J, Liang L, Song H, Kong Y, Ma R, Hou Y, Zhao J, Liu J, He N, Zhang Y: A method for hand-foot-mouth disease prediction using GeoDetector and LSTM model in Guangxi, China. Sci Rep 2019, 9(1):17928.
Norris D: Short-term memory and long-term memory are still different. PSYCHOL BULL 2017, 143(9):992–1009.
Kirbas I, Sozen A, Tuncer AD, Kazancioglu FS: Comparative analysis and forecasting of COVID-19 cases in various European countries with ARIMA, NARNN and LSTM approaches. Chaos Solitons Fractals 2020, 138:110015.
JuHyok U, Lu P, Kim C, Ryu U, Pak K: A new LSTM based reversal point prediction method using upward/downward reversal point feature sets. CHAOS SOLITON FRACT 2020, 132.
Duong TA, Van SP, Dang TD, Hoang LP: Downscaling rainfall using deep learning long short-term memory and feedforward neural network. INT J CLIMATOL 2019, 39(10):4170–4188.
Wang G, Wei W, Jiang J, Ning C, Chen H, Huang J, Liang B, Zang N, Liao Y, Chen R et al: Application of a long short-term memory neural network: a burgeoning method of deep learning in forecasting HIV incidence in Guangxi, China. EPIDEMIOL INFECT 2019, 147:e194.
Gu J, Liang L, Song H, Kong Y, Ma R, Hou Y, Zhao J, Liu J, He N, Zhang Y: A method for hand-foot-mouth disease prediction using GeoDetector and LSTM model in Guangxi, China. Sci Rep 2019, 9(1):17928.
Anand S, Hanson K: Disability-adjusted life years: a critical review. J HEALTH ECON 1997, 16(6):685–702.
Murray CJL, Lopez AD: Measuring the Global Burden of Disease. NEW ENGL J MED 2013, 369(5):448–457.
Murray CJ, Vos T, Lozano R, Naghavi M, Flaxman AD, Michaud C, Ezzati M, Shibuya K, Salomon JA, Abdalla S et al: Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study 2010. LANCET 2012, 380(9859):2197–2223.
Wang YW, Shen ZZ, Jiang Y: Comparison of autoregressive integrated moving average model and generalised regression neural network model for prediction of haemorrhagic fever with renal syndrome in China: a time-series study. BMJ OPEN 2019, 9(6):e25773.
Yang J, Li L, Shi Y, Xie X: An ARIMA Model With Adaptive Orders for Predicting Blood Glucose Concentrations and Hypoglycemia. IEEE J Biomed Health Inform 2019, 23(3):1251–1260.
Chaurasia V, Pal S: COVID-19 Pandemic: ARIMA and Regression Model-Based Worldwide Death Cases Predictions. SN Comput Sci 2020, 1(5):288.
Maragatham G, Devi S: LSTM Model for Prediction of Heart Failure in Big Data. J MED SYST 2019, 43(5):111.
Yu Y, Si X, Hu C, Zhang J: A Review of Recurrent Neural Networks: LSTM Cells and Network Architectures. NEURAL COMPUT 2019, 31(7):1235–1270.
Wang G, Wei W, Jiang J, Ning C, Chen H, Huang J, Liang B, Zang N, Liao Y, Chen R et al: Application of a long short-term memory neural network: a burgeoning method of deep learning in forecasting HIV incidence in Guangxi, China. EPIDEMIOL INFECT 2019, 147:e194.
Gu J, Liang L, Song H, Kong Y, Ma R, Hou Y, Zhao J, Liu J, He N, Zhang Y: A method for hand-foot-mouth disease prediction using GeoDetector and LSTM model in Guangxi, China. Sci Rep 2019, 9(1):17928.
Kirbas I, Tuncer AD, Sirin C, Usta H: Modeling and developing a smart interface for various drying methods of pomelo fruit (Citrus maxima) peel using machine learning approaches. COMPUT ELECTRON AGR 2019, 165.
Alim M, Ye GH, Guan P, Huang DS, Zhou BS, Wu W: Comparison of ARIMA model and XGBoost model for prediction of human brucellosis in mainland China: a time-series study. BMJ OPEN 2020, 10(12):e39676.
Adeyinka DA, Muhajarine N: Time series prediction of under-five mortality rates for Nigeria: comparative analysis of artificial neural networks, Holt-Winters exponential smoothing and autoregressive integrated moving average models. BMC MED RES METHODOL 2020, 20(1):292.
Sahin U, Sahin T: Forecasting the cumulative number of confirmed cases of COVID-19 in Italy, UK and USA using fractional nonlinear grey Bernoulli model. Chaos Solitons Fractals 2020, 138:109948.
Gu J, Liang L, Song H, Kong Y, Ma R, Hou Y, Zhao J, Liu J, He N, Zhang Y: A method for hand-foot-mouth disease prediction using GeoDetector and LSTM model in Guangxi, China. Sci Rep 2019, 9(1):17928.

No competing interests reported.

PDF

Reviews received at journal
22 Apr, 2022
Reviewers agreed at journal
12 Apr, 2022
Reviewers invited by journal
12 Apr, 2022
Editor assigned by journal
12 Apr, 2022
Editor invited by journal
12 Apr, 2022
Submission checks completed at journal
12 Apr, 2022
First submitted to journal
02 Mar, 2022

You are reading this latest preprint version

Comparison of ARIMA model and LSTM model in predicting disease burden of occupational pneumoconiosis in Tianjin, China

Status:

Version 1

Abstract

Background

Methods

Restults:

Conclusion

Figures

Background

Methods

Data source

DALY Calculation

ARIMA model

LSTM model

Model Comparison

MAE=\(\frac{1}{n}{\sum }_{i=1}^{n}\)|\({\widehat{y}}_{i}-{y}_{i}\)| (4)

RMSE=\(\sqrt{\frac{1}{n}{\sum }_{i=1}^{n}({\widehat{y}}_{i}-{y}_{i})2}\) (5)

MSE=\(\frac{1}{n}{\sum }_{i=1}^{n}\)(\({\widehat{y}}_{i}-{y}_{i}\))2 (6)

MAPE=\(\frac{100\%}{n}{\sum }_{i=1}^{n}\)|\(\frac{{\widehat{y}}_{i}-{y}_{i}}{{y}_{i}}\)| (7)

Statistical Analysis

Results

Trend of pneumoconiosis burden in Tianjin

Fitting models with ARIMA

Sequence stabilization

Model identification and screening

Model test and prediction

Fitting models with LSTM

Data normalization processing

Establishment of LSTM Model

Prediction results

Model comparison

Prediction of DALY for future pneumoconiosis in Tianjin

Discussion

Conclusion

Abbreviations

Declarations

References

Competing Interests

Status:

Version 1

MSE=\(\frac{1}{n}{\sum }_{i=1}^{n}\)(\({\widehat{y}}_{i}-{y}_{i}\))² (6)