DATA DESCRIPTION & ASSUMPTIONS

In this section, the group developed the simulation model of the hospital ED by mapping the path of patients (entities) through their COVID-19 hospital stay. The entities included in the simulation possess certain attributes characterizing their treatment and tenure in the ED. They also seize resources such as doctors, nurses, beds, and equipment to aid in their recovery.

To create the simulation model, the group utilized anonymous data from an unidentified US Hospital. The data is presented in two datasets. The first dataset contains the patient hospitalization data. Each piece of data provides a patient id (denoted as PAT_ID), an encounter id (PAT_ENC_CSN_ID), the patient’s admission time into the hospital (HOSP_ADMSN_ TIME), along with the severity of the patient’s health condition which ranges from 0 (not severe) to 1 (extremely severe). The patient’s living status (LIVING_STAT) denotes whether the patient dies as a result of COVID-19 or if they recover. If the patient is diseased, the date of their death is recorded (DEATH_DATE). This date may or may not be equivalent to the end date of the patient’s hospitalization (END_DATE) which is recorded for both living and deceased patients. The END_DATE of a patient who dies may be later than the DEATH_DATE since there is a slight delay in which the deceased patient is removed from the hospital after death.

The second dataset provides more specific information for patients who stay in the ICU. Each patient who stays in the ICU has their patient id and encounter id recorded, along with the date when the patient stayed in the ICU (DATE). DATE provides both the entry and exit times of the ICU for each patient.

The group’s initial step was to prepare the data for use in the model by performing some data cleaning. First, the data provided included patients who entered the system, exited the system, and then re-entered the system. This is most likely due to patient hospital transfers or could be from patients who appeared to be mostly recovered, and then took a turn for the worst, requiring re-hospitalization. In the dataset, these patients will have the same PAT_ID for both hospital visits but will have a different PAT_ENC_CSN_ID. To deal with this scenario, the group assumed that each new PAT_ENC_CSN_ID represents an entirely new patient since this ID is unique for each patient who enters the system. In this case, both hospitalizations of the same patient are treated in the model as seperate hospital visits by different patients.

In order to more effectively utilize the ICU data from the second dataset, the group merged this dataset with the hospitalization data. This way, it was easier to observe which patients were hospitalized in the ICU, and which remained in regular impatient beds. Since the dates in the ICU dataset only contain the day of patient arrival/departure and not the specific time, the group decided to ignore the time part of the date in the hospitalization data. For example, patient 15134993 was admitted on April 1 2020 at 16:41:00 (shown in the dataset as 01APR2020: 16:41:00). The group removed the time portion of this entry so the hospital admission time now shows only 01APR2020.

Now that it was determined which patients went to the ICU first and which patients went to a regular bed first, the patients from these two different cases were able to be sorted. After these patients entered the system, they passed through a variety of routes before they exited the system. They may have stayed in the bed of their arrival for the duration of their hospital stay, or they might have moved multiple times between an ICU bed and regular bed. To simplify the model, the group assumed that only one transfer at most occurs for patients. This means that if a patient is transferred from a regular to ICU bed, they will stay in the ICU bed and not transfer back to a regular bed. With this information, several probabilities were calculated. First, the probability of patient arrival to the ICU first or to a regular bed first was calculated. Then, the probabilities of patient transfers between different types of beds were calculated. Finally, the probability that a patient exits from the system due to death or recovery was calculated.

Finally, the group used the dataset to plot the arrival time of the patients. Afterwards, the group decided to use a non-stationary poisson process to calculate the different lambda (arrival times) values for each period of time (day). Each arrival was assumed to be independent and a separate patient. A non-stationary poisson process was chosen because the arrivals of patients into the hospital are not constant and are time series dependent. This method allows the arrival of patients in the simulation to closely represent the reality of patient arrivals into the hospital.

The final assumption made was that there was no balking in the system. In the model, once customers arrive, they wait until they are able to be served. However, in a real system, upon seeing a long queue, some customers would decide not to enter the queue and would instead leave the system.

The only piece of data from the dataset provided that was unused is the patient’s severity. This is an attribute that the group is able to use in future iterations of the model to increase the model’s complexity.