ICU Bed Capacity Analysis with Transformer-Based Length of Stay Prediction and Erlang Loss Formula

This paper introduces a new method for predicting the length of stay in the ICU using a transformer-based architecture. The predicted results are used for analyzing ICU bed capacity with the Erlang loss formula. The prediction architecture comprises two transformer layers, one global average pooling layer, and two fully connected layers. The proposed model classifies the length of stay into four classes depending on how long patients are admitted to the ICU based on patient data. In addition, the model estimates the length of stay in days of each patient. We then utilize the estimated length of stay to analyze ICU resources with the Erlang loss model. Given ICU blocking probabilities, the required ICU bed numbers are estimated with Erlang-B. Patient data from the MIMIC-III dataset, including various vital signs and patient health conditions, and the actual statistics from the Pediatric ICU at King Chulalongkorn Memorial Hospital are deployed to evaluate the proposed method. Experimental results show that the transformed-based architecture surpasses previous works regarding length of stay prediction accuracy. In addition, our analysis of ICU resources provides close results to those from the actual statistics from the Pediatric ICU at King Chulalongkorn Memorial Hospital.