Predicting health care facility stay duration: a machine learning approach

The COVID 19 pandemic revealed shortcomings in healthcare, particularly concerning bed occupancy and resource allocation. During the Delta variant wave, it was highlighted how much improvement is needed in management strategies. One promising solution is the prediction of inpatient Length of Stay. Accurate predictions can enhance efficiency, reduce infection risks, lower mortality rates and decrease bed occupancy. This research proposes a predictive model using Random Forest Regression to accurately forecast hospital length of stay, aiming to enhance resource management and patient care. We utilized a 2010 inpatient dataset from the New York Department of Health and conducted thorough data preprocessing, including cleaning, handling missing values, and numerical encoding of categorical variables for regression. Additionally, we experimented with three database variations: one with targeted and frequency encoding, another using synthetic minority oversampling technique for handling imbalances, and a third applying synthetic minority oversampling technique for regression with gaussian noise for continuous variables. Each database was tested with and without scaling using four different scalers. The objective was to achieve a mean absolute error below the industry standard of 6.5, prioritizing unbiased metrics. Our results indicate that the final model achieved a 2.93 mean absolute error on the normal database, demonstrating its effectiveness in predicting length of stay. The study underlined the potential of machine learning in accurately predicting the Length of Stay in hospitals and the possibility of a more accurate model of the industry standard. Further advancements could be made to the models with more balanced datasets and a user-friendly interface for hospital staff usage.