Time Series Analysis Of The Admitted Patients With Chronic Pancreatitis

Objectives: By analyzing the monthly time series of admitted patients with chronic pancreatitis in our hospital in recent years,our study explored the periodicity and seasonality of the admission of patients with chronic pancreatitis,and investigated the methods to predict the admission of patients with chronic pancreatitis in the future.Method:The time series of the number of hospitalized patients with chronic pancreatitis in recent 143 months was generated based on the data of the Second Hospital of Hebei Medical University from December 2006 to October 2018.First,the basic characteristics of the patients with chronic pancreatitis in our hospital were analyzed.Secondly,the time series of the first 123 months were subjected to differentiation,order determination,and then be fitted to an optimized ARIMA model though the function of “auto.arima”.The optimized model was utilized to forecast the number of hospitalized patients for the next 20 months.LSTM model was also used to train the data of the first 123 months and test the data of the last 20 months.Finally,the predicted results of the two models were compared with the real values of the last 20 mouths.Data processing and modeling were carried out in R language.Among them,the modeling of ARIMA mainly relied on the R language packages “forcast” and “tseries”,while LSTM relied on “keras” which is a high-level deep learning framework developed by Google.Results:1.Firstly,the basic characteristics and the seasonality of the admitted patients with chronic pancreatitis in our hospital were analyzed.The number of admitted patients with chronic pancreatitis was steadily increasing over the last decade.The majority of admitted patients were male in middle or old ages.Most of the inpatients were the first time or the second time admitted to our hospital.Approximately 10% of the data could attribute to seasonal factors,and the peak of admission was in the third quarter.2.Two methods,ARIMA and LSTM,were utilized to training a model to predict the admitted patients of last 20 mouths.The RMSE and MAE were calculated to evaluated the predictive strength of models.It is shown that the prediction based on LSTM is with smaller deviation from the real values than the prediction based on ARIMA.However,the method of ARIMA is more readable as the parameter settings is concise during the modeling of ARIMA.Conclusion:1.The number of inpatients with chronic pancreatitis could be described by a times series with seasonality.2.LSTM model based on recurrent neural network could accurately predict the number of patients with chronic pancreatitis and the overall trend,while the predictive strength of ARIMA is weaker than LSTM.But,complicated parameters need to be trained during the modeling process of LSTM,which may limit its applications in time series data in medical fields.