Offline Evaluation of Predicting Hospital Acceptance Using Machine Learning

Abstract

　Objective: To explore the potential application of machine learning in hospital selection by emergency medical services (EMS) through the development of offline evaluation methods. Methods: We examined two approaches: 1) predicting acceptance at hospitals selected by EMS personnel, and 2) hospital selection solely based on machine learning. Results: In the scenario where EMS personnel alone faced a 29.96% rejection rate in total inquiries, approach 1) reduced the total number of inquiries by 5.18% to 19.0%. However, models with higher reduction rates showed an increase in cases with uncertain acceptance status. For such cases, we presented a method to calculate how much assumed acceptance probability would be needed to outperform EMS personnel alone. Approach 2) achieved a 20.55% to 21.44% reduction in inquiries. We identified that this method also produced cases with uncertain acceptance status and presented similar countermeasures. Discussion: For approach 1), it is advisable to start implementation with models that have a lower proportion of uncertain inquiries. For approach 2), cases with uncertain acceptance status tend to be among those with higher acceptance probabilities as output by the model, suggesting that actual acceptance rates for these cases might be high.