Quantity Prediction Of Dermatology Outpatients Based On MBGA Feature Selection And LSTM

With the rapid development of the Internet,Internet of Things,Artificial Intelligence and big data technology,there are increasing number of hospitals applying these new technologies to their daily operations.These emerging technologies are used in doctors'scheduling,drug procurement,medical device distribution,etc.,and can directly improve the hospital's operational functions and medical effects.For hospitals,the number of patients per day is an important index for hospital decision-makers to make correct decisions.Taking dermatology department as an example,a more reasonable patient prediction method can save medical resources and provide patients with a good experience.In the literature review of outpatient population prediction,most researchers build a time series model by simulating the number of historical outpatients,without considering the impact of external factors such as climate on the occurrence of diseases.Based on previous studies,this paper considers the regularity of the number of patients with various skin diseases under the influence of multiple factors(such as climate and environment).It is precisely because there are few studies on feature selection before training prediction models,but in the field of data mining,this is an important preprocessing operation.In this paper,a novel hybrid method is proposed to predict the number of patients with different skin diseases by combining new feature selection methods with in-depth learning methods.The research contents are as follows:1.Firstly,without considering the characteristics of climate,this study only observes the changing trend of the number of outpatients with various skin diseases according to the historical data of dermatology clinics.After data difference,order determination and model evaluation,the Autoregressive Integrated Moving Average model(ARIMA)was established to observe the ability of the model to predict the number of patients.2.Then the data are preprocessed by climatic data,air quality data,date,and the features I constructed,after that integrated into a new feature set.After adding external factors,aiming at the high-dimensional data set of small samples,in order to prevent the over-fitting of cyclic neural network model in training samples,this paper modified binary genetic algorithm(MBGA)for feature dimensionality reduction and embedding it into LSTM model for prediction,and observes whether the prediction accuracy is improved and what factors affect the number of patients with dermatosis.By introducing LSTM model,the influence of multiple parallel time series variables on the output can be considered,therefore,to improve the accuracy of prediction.MBGA combines filtering and wrapping feature selection methods to accelerate iteration while improving accuracy.In the design of fitness,multi-objective optimization is realized.Which is minimizing the average absolute percentage error(MAPE)and the number of selected features.Besides,an optimal feature ensemble step is added to improve the robustness of the selected feature subset.After iteration of MBGA-LS model,higher prediction accuracy is achieved according to fewer features.3.Finally,the performance of MBGA-LSTM model is evaluated from two aspects.first aspect is to compare various feature selection methods.Comparing MBGA-LS with GA-LSTM,PCA-LSTM and prediction models without feature processing.It is found that MBGA improves the iteration speed while ensuring the accuracy.following aspect is to compare MBGA with various prediction models.In order to prove the superiority of LSTM network combined with MBGA,MBGA-LSTM is compared with MBGA-MLP,MBGA-SVR and MBGA-LR.Which is found MBGA-LSTM has the most stable prediction ability and the strongest comprehensive performance among all the combinations involved.