Performance Evaluation of Artificial Intelligence Methods Predicting Annual Number of Patients in Hospitals

Abstract

Background: Multiple artificial intelligence (AI) methods are being applied to estimate health care sectors' capacities. This study evaluated AI techniques' predictive performance for determining the annual number of patients in hospitals. Methods:The units of analysis were individual hospitals. This study used a dataset consisting data from 708 hospitals in 2021. Training and test datasets were divided by a ratio of 8:2. The two dependent variables were total length of inpatient stay (TLOS) and total number of outpatient visits (TNOV). Four machine learning techniques were used: linear regressor (LR), random forest (RF), gradient boosting model, and extreme gradient boost. Model performance was evaluated with coefficient of determination (R 2 ) in addition to mean squared error (MSE), mean absolute error (MAE), and root mean square error (RMSE). The study used Python version 3.7.0. Results:The best-fit model for predicting TLOS (R 2 =0.730) and TNOV (R 2 =0.707) was RF and LR, respectively. Both RF and LR also had the lowest MSE, MAE, and RMSE scores for inpatient and outpatient predictions. The most remarkable factor associated with a good prediction of the inpatient sector was number of beds followed by number of nurses. In contrast, the best predictive factor for the outpatient sector was number of doctors followed by number of local households. Conclusion:This study confirmed that AI methods are successful at predicting hospitals' annual patient loads. Among them, the RF and LR models showed the best performance in predicting both inpatient and outpatient annual loads, respectively. This study proposes that the AI tools utilized in this study can accurately predict future medical demand.