Machine Learning Methods for Prediction of COVID-19 Patient Length of Stay: Using Texas PUDF Data

Accurately predicting the length of stay (LOS) has significant benefits for hospitals’ resource utilization. It’s especially important for COVID-19 patients since it is a highly contagious disease, and it places significant strain on healthcare systems worldwide. In this paper, we used the Texas Public Use Data File (PUDF) inpatient discharge record, which is a state-level dataset that contains data from over 600 hospitals, to make predictions regarding LOS of COVID-19 patients. Our analysis included 224,152 (n) records covering a time span of six quarters from Q4 2020 to Q1 2022. In addition, seven features are chosen to predict LOS - age group, sex code, race, first payment source, type of admission, illness severity, and risk mortality. Various machine learning techniques were employed, including Gradient Boosting (GB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF). Using regression models, we identified the best parameters for each model and predicted the numerical LOS for each quarter. Additionally, the Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-Squared (R2) were calculated to evaluate the performance of each prediction. The results revealed notable performance differences among models. SVM performs the least MAE for 6 quarters, and GB performs the highest R-squared for 6 quarters. It also showed varying performance across different quarters, possibly due to dissimilarities in the data distribution. Q4 2020 and Q2 2021 has a relatively better performance, and the data is relatively concentrated compared to others. Q4 2021 has the least performance with a highly dispersed data distribution.