Modelling hospital operations: insight from using data from paper registries in the obstetrics ward at a hospital in Addis Ababa, Ethiopia

In the Ethiopian health system, operations management techniques have been underutilised. Although previous research has outlined limitations of paper-based patient records, few studies have examined their potential utility for improving management of hospital operations. In this paper, we used data collected from paper registries in an Ethiopian obstetrics ward at Addis Ababa’s Tikur Anbessa Specialized Hospital, Ethiopia’s largest university hospital, to model the ward’s operations. First, we attempted to identify predictors of lengthy stays and readmissions among women giving birth: few predictors were deemed significant. Second, time series methods for demand forecasting were applied to the data and evaluated with several error metrics, and these forecasts were improvements over baseline methods. We conclude with recommendations on how the obstetrics ward could incorporate our modelling approaches into their daily operations.

significant association with the risk of an extended LOS. Older patients seem to remain in the ward longer than younger ones; a one-year increase in age was associated with a 6% (95% CI: 1.02-1.10) increase in the risk of an extended LOS. Patients giving birth to more than one child were over 5 (1. 44-18.39) times more likely to have an extended LOS compared to patients giving birth to one child. Similarly, patients with pre-eclampsia had an odds ratio of 4.47 (1.42-14.03) for an extended LOS. The model showed no significant difference in risk of extended LOS among patients based on delivery method, daily volume, or region of residence. We found no significant results in our analysis of the risk of readmission (Table A1).
Our analysis showed that patients with multiple births at TASH had an OR for extended stays of 5.14. Campbell et al. found that having twins or triplets was associated with an average increase in LOS of close to two days.    Table A2 for details on model formulations.

Naïve forecast
The naïve, or one-step, forecast predicts the value Ŷ at time t + n to be the actual value of Y in the previous period t. For weekly forecasts, n = 1. For daily forecasts, n = 1 and 7, i.e. steps of one day and one week.
The moving average predicts Ŷ in the period t + 1 to be the average of actual values during the previous N periods, where N is called the window size. For the weekly forecasts, we tested window sizes of 2, 3, and 4 weeks. For daily forecasts, we tested window sizes of 3, 5, and 7 days.

Exponentially weighted moving average (EWMA)
The exponentially weighted moving average (EWMA) discounts past values of Y and Ŷ according to a scaling factor α. The scaling factor α can be written as a function of a window with N periods.
For daily forecasts, we tested window sizes of 3, 5, and 7 days. For the weekly forecasts, we tested window sizes of 2, 3, and 4 weeks.

Daily historical mean
The daily historical mean model calculates mean admissions 8 by day D of the week in the training period and applies those means to the testing period. The predicted number of admissions on a Wednesday, for example, is the average number of admissions on Wednesdays in the training set. The training period for daily admissions was August 2015. The training period for weekly admissions was August 2015 to January 2016. " #$% = 8 9

Historical mean for weekdays and weekends
The model predicts Ŷ on day D as the mean daily admissions 8 : during either weekdays or weekends, depending on D, in the training period. We considered Monday through Friday to be weekdays and Saturday and Sunday to be weekends. The training period for daily admissions was August 2015. The training period for weekly admissions was August 2015 to January 2016.