Abstract: BACKGROUND: The length of hospital stay after total knee arthroplasty is closely related to the prognosis of patients, but the related factors that affect the length of hospital stay have not been studied in depth.  
OBJECTIVE: The prediction model of length of stay after total knee arthroplasty was established based on clinical data. Seven machine learning algorithms were used to construct the model, evaluate the effectiveness of different algorithms, and obtain the most relevant influencing factors of length of stay.
METHODS:  Through the hospital medical record system, a total of 777 patients who underwent total knee arthroplasty that met the inclusion criteria from January 2012 to December 2019 were collected. The patient’s clinical data and past medical history were entered in detail. Seven kinds of machine learning, such as logistic regression, multiple adaptive regression, K-nearest neighbor, support vector machine, random forest, extreme gradient extraction algorithm, and artificial neural network, were used to build an algorithm model, and the 10-fold cross-validation method is used to verify the effectiveness of the model. The area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, accuracy and F1 score of the seven models were calculated and compared. The importance of the predictive variables of the artificial neural network model, the neural network architecture and draw heat maps related to the traits were evaluated.  
RESULTS AND CONCLUSION: (1) We included a total of 777 samples, including 618 patients who were hospitalized for less than or equal to 6 days and 159 patients who were hospitalized for more than 6 days. (2) There were significant differences in age, preoperative hemoglobin, operation method, diabetes history, ischemic heart disease history, cerebrovascular disease history and blood transfusion between the two groups (P < 0.05). (3) Logistic regression, multiple adaptive regression, K-nearest neighbor, support vector machine, random forest, extreme gradient algorithm and artificial neural network area under the receiver operating characteristic curve were 0.770, 0.778, 0.609, 0.570, 0.594, 0.586, and 0.903 in sequence. The prediction efficiency of artificial neural network was the best. Simultaneously, through the comparison of accuracy, sensitivity, specificity, precision and F1 score, it is found that the artificial neural network performs best, followed by logistic regression and multiple adaptive regression algorithms. (4) In the artificial neural network model, the length and age of surgery were the most important among the predictors, leading the other predictors, and the history of heart failure, cardiovascular disease, ischemic heart disease, surgical methods, hemoglobin, blood transfusion, insulin use, history of diabetes, and obstructive sleep apnea were less important than the first two, but they were still highly correlated predictors. The neural network architecture also confirms the importance of these factors. (5) It is concluded that the three machine learning algorithms of artificial neural network, logistic regression and multiple adaptive regression algorithm can be used to predict the length of hospitalization after total knee arthroplasty, but the prediction effect of artificial neural network was more than that of logistic regression and meta-adaptive regression algorithm accurate. The length of surgery, age, history of heart failure, cardiovascular disease, ischemic heart disease, surgical method and hemoglobin are closely related to the length of hospitalization in the artificial neural network model. Simultaneously, using the neural network model can personally predict the patient. The artificial neural network prediction model has a high recognition efficiency, which helps to improve the utilization rate of hospital beds and better plan the length of hospital stay.

Key words: bone, knee, joint, machine learning, arthroplasty, length of hospital stay, artificial neural network, prediction model