Discharge Prediction for Critical Patients Using Machine-Learning Technology

Abstract

To predict four forms of discharge—death, home, nursing facility, and rehabilitation—machine learning models were developed. These models were trained on 115,248 distinct ICU admissions, and they were assessed. To lessen class disparity, we used synthetic minority oversampling techniques. Hierarchical and ensemble classifiers were used to examine the impact of an unequal testing set on the performance of our prediction models. With a 90% area under the receiver operating characteristic curve (recall 71%, F1: 70%), the XGBoost outperformed earlier tests in terms of discriminating performance. Our research demonstrates that the variables included in the proposed approach, known as the k-SMO (K-Sequential Minimal Optimization) model for assessing patient illness severity, are more accurate predictors of hospital release location than the current work score alone. Clinical decision support systems that incorporate these models may benefit patients, employees, and ICU staff by enabling them to begin disposition planning as soon as practical throughout their stay. The length of a hospital stay and its effects have a big effect on people and the economy. The present effort has focused on forecasting that crucial parameter as a consequence. Study groups have achieved successful prediction rates, although only very seldom revealing predictive patterns. The length of stay (LOS) forecasting model we present in this research, along with a study of trends and patterns that show how LOS differs amongst hospital departments and specialties, are also included. We also look into whether patient data might help hospital departments predict LOS more accurately. It is helpful to know where to concentrate attention while expecting LOS since the findings for these known patterns are often up to 21.61% better in terms of estimating error rates and up to 23.83% higher in terms of projecting success rates. Several trends appeared after the estimation of prediction rates. For women who were admitted to the emergency room for enteritis-related problems, for instance, these patterns improved prediction accuracy. Overall, the results show that the calculated error rates for these recognized patterns are up to 21.61% lower and the success rates for the number of predictions are up to 23.83% higher, which is useful for figuring out where to put the most effort when predicting LOS.