Existing System & Flaws

Kim et al. [12] predicted the hospitalization of patients visiting the ED and showed which characteristics of the patients influenced their likelihood ofhospitalization. Their predictive model was mainly analyzed based on accuracy and the area under the ROC curve (AUC). They found that the older the patient and the more urgent their condition, the more likely they were to be hospitalized. Lucke et al. [13] divided 21,287 ED patients into two groups (>70 years and <70 years old) and predicted hospitalization using LR. They evaluated patients' hospitalization predictions using indices like AUC and positive prediction values. Their study demonstrated that predictive models could help identify patients who were more likely to be hospitalized using readily available information, such as their vital signs. Graham et al. [11] used three algorithms, namely LR, the gradient boosting model (GBM), and decision tree (DT), to predict the hospitalization for ED patients and analyzed 107,545 patient data. They suggested that when choosing a predictive model, simplicity and interpretation efficiency took precedence over the model's performance. Some studies have considered using neural networks in addition to regression and ensemble-based classifications for hospitalization predictions. Araz et al. [14] performed hospitalization predictions based on LR, DT, support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), and artificial neural network (ANN) models using data from 118,005 patients. Among the predictive models, XGBoost showed the highest AUC. Hong et al. [8] analyzed LR, XGBoost, and deep neural networks (DNNs). Based on 560,486 patient visits, they analyzed three groups of data: patient severity classification data, clinical data from previous visits, and all the available data from previous and current visits. In this analysis, XGBoost and DNN displayed good AUC values when predicting ED patient hospitalizations. Golmohammadi [15] presented hospitalization predictions using LR, ANN, and a statistical method that patterned the similarity of patient characteristics to predict hospitalization. He showed that the overall accuracy of the three models was greater than 80%. Dinh et al. [17] limited the targets of analysis to adult patients aged 16 years or older and included 860,832 patient data in the analysis. LR was used to predict hospitaliza-tions to improve the patient flow and aid clinical decision- making in the ED. LR was interpreted based on AUC. Their study showed that accurate hospitalization predictions for ED patients could be made using initially available patient information, such as age, mode of arrival, and time of arrival. Fenn et al. [18] constructed a predictive model using Light- GBM to divide the likelihood of the hospitalization into four stages: low, medium, high, and very high. A total of 468,167 patient data points were used. Medical personnel could respond flexibly to patients' follow-up processes by dividing them into several categories according to their like- lihood of hospitalization. This constructed predictive model was measured based on AUC. Goto et al. [19] studied hospitalization predictions for children who visited the ED. Using data from 52,037 children, they used lasso regression, RF, XGBoost, and DNN to predict two clinical outcomes: critical care or hospitalization. These models were evaluated for their sensitivity and specificity; DNN was the best predictor of hospitalization in children. Horng et al. [20] used data from 230,936 ED patients to predict their dis- eases. In addition to circulated data, such as patients' vital signs, they used free-text information on important symptoms to identify patients with sepsis. They employed an SVM to predict sepsis and assessed it based on AUC. The SVM achieved high performance using data routinely avail- able during triage (e.g., reasons for visits and vital signs). Ram et al. [21] conducted a study to predict the number of daily ED visits of asthmatics using Twitter and Google data collected from various regions. They showed that a predictive model using data over a short period of three months could predict the ED visits of asthmatics in near real-time. The DT and ANN models that were used predicted the number of ED visits for asthmatics in daily low-, medium-, and high-volume categories. The models were evaluated in terms of their AUC and precision values; the predictive accuracy of the medium- volume category obtained using the hybrid ANN and DT models was the highest. Barack-Corren et al. [7] studied the prediction of hospitalization for pediatric patients who visited the ED. A total of 59,033 patient data points were used, and the predictions were tested using the data that were available within 10, 30, and 60 min after the patients arrived at the ED. The predictions were made using a hybrid model that combined Naive Bayes and LR. They estimated the potential effects of hospitalization predictions for ED patients. From the perspective of the ED patients' flow, they derived the effects of the time saved in the ED and the time costs in the inpatient ward (i.e., the total time during which empty beds were held for ED patients in the inpatient ward). Disadvantages ? There is no an effective MACHINE LEARNING ALGORITHMS used in the exising system. ? The system not implemented MODEL FITTING AND EVALUATION in the datasets.