Developing data science approaches to improve paediatric critical care patient flow and its related health economic benefits

Abstract

In recent years, a worsening bed crisis in Paediatric Critical Care Units (PCCU) has been observed in Scotland. There are many reasons for this, including increasing seasonal respiratory virus cases and increasing patient complexity and numbers. Patient flow is the movement of patients between different hospital units during their admission. It is an essential factor to consider when maximising the resource efficiency of a healthcare system. A significant bottleneck to patient flow through PCCU is bed capacity, understanding this further could help alleviate this bed crisis. This thesis has investigated applying data science techniques to two types of routinely collected data to understand or improve bed capacity in the PCCU within the children’s hospital in Edinburgh, Scotland.

Novel bed availability forecasting models were developed using 8 years of routinely collected bed management data. These models were developed to forecast both the bed availability and whether one or more beds will be available up to 14 days into the future in all units and wards in the hospital. Various time series forecasting methods were assessed with different temporal and resource-related covariates. Overall, simpler linear models outperformed or equalled the performance of traditional statistical autoregressive models and state-of-the-art deep learning models. It was observed that the forecasting performance decreased as the forecasting horizon increased to 14 days, where the performance of the models converged towards the baseline random walk model. Feature importance indicated that the models focussed mainly on resource information related to the unit/ward being modelled and temporal features.

For the first time, routinely collected clinical grade minute-by-minute physiology data from a UK multi-centre dataset was used to predict whether Traumatic Brain Injury (TBI) patient PCCU length of stay was longer than four days. Predicting patient LoS could aid key stakeholders in planning future bed availability. Machine learning was applied directly to the physiological time series using deep learning methods. Additionally, four feature extraction techniques applied to the physiological time series were assessed, including (i) TBI clinical feature extraction, (ii) automated time series summarisation, (iii) physiological pattern mining and (iv) automated physiology clustering. Model performance was assessed using nested-cross validation. Only the TBI clinical feature extraction algorithm and automatic time series summarisation techniques performed better than the baseline. In contrast, the other methods showed significant signs of overfitting.

For the first time, a hybrid discrete event and agent-based simulation of a PCCU has been developed. The simulation was designed to digitally replicate the RHSCE with the ability to test various staffing and patient-related scenarios, including a change in bedside nurse skill set and complexity, as well as changes in patient numbers. The discrete event component of the simulation replicates the management of the hospital, including staff scheduling and assignment of resources, and the agents in the simulation include nurses, clinicians, patients, and beds. It was demonstrated that increased patient complexity, and a reduction of nurse skill level resulted in longer waiting times for care due to increased pressure on experienced nurses, while increased admissions to the general wards and PCCU caused longer PCCU LoS due to delayed discharges.

An application of the bed availability forecasting models to dynamically schedule bedside nurses to reduce PCCU bedside nurse staffing resource demand and evaluate the total number of bedside nurses was investigated. The number of nurses was dynamically scheduled over a 14 day window using the forecasted bed demand. Nurse scheduling was conducted using mixed integer optimisation over the scheduling window. In a first-of-its-kind study, the methodology was robustly tested using discrete event simulation to account for significant random events. Results showed that by using model-based scheduling methods, it was possible to reduce nurse staffing resource demand. It was, however, identified that there was a higher risk of the unit being understaffed when model predictions were incorrect, especially over longer forecasting horizons. Implementation of model-based scheduling was only possible when using flexible nurse staff systems such as on-call nurses.

Overall, this thesis has demonstrated how routinely collected data that are currently underutilised can be used to better understand and improve patient flow through PCCU in Scotland.