Prediction of length of stay for primary THA/TKA patients using machine learning in OCON Orthopedische Kliniek.

OCON experiences problems resulting from the difficulty to predict the postoperative length of stay for primary total knee/hip arthroplasty patients, such as insecurity for patients or suboptimal capacity utilisation. We make a selection of patient-related features based on literature and extract patient data from OCON's database. We select four feature selection methods and eight machine learning methods. In total, we train 34 machine learning models for each surgery type (knee/hip). We make use of cross-validation and apply data scaling. We select the best performing models based on their ROC curve AUC values as well as the number of features they require. We select the most promising models, which have an AUC around 0.72 and 0.76, and the number of features required ranges from 8 to 22. Despite a small data of 5400 surgeries for both hip and knee surgeries, the best models achieved AUC values comparable to literature. We create a prediction tool where users can select the surgery type and machine learning method. After inputting values for the corresponding input features, the tool makes a prediction on whether the patient is expected to experience a length of stay of at least 3 days.