Keywords

1 Introduction

Health-care managers are currently under constant pressure to control rapidly escalating expenses, while still responding to growing demands for both high-class patient service levels and medical treatment. Resolving such challenges requires a consistent understanding of health-care systems, which can be an overwhelming task, given the high levels of uncertainty and interdependence. Overcrowding in emergency departments (EDs) has become a significant international crisis that negatively affects patient safety, quality of care, and patient satisfaction [5, 7, 8, 16]. The research on healthcare system efficiency improvement has attracted much attention, in which Discrete Event Simulation (DES) based approach appears to be a dominant tool. Efforts to develop simulation models have advanced since the late 1980s when simulation was used to investigate the impact of key resources on waiting times and patient throughputs [15], and it has since been used to study the effect of a wide range of health interventions on health-care processes’ performance [3, 4, 11, 12].

In spite of these advances in analyzing and improving healthcare processes, new approaches are still needed for two reasons. First, the majority of reported studies are not flexible and cannot be adapted or reused [9]. Second, due to the complexity of healthcare process, it is challenging to get a precise view of the patient flow in the health system with enough details. Therefore, the majority of DES studies develop the process model from observations which is not only unrealistic but also time-consuming. Process models are the core component of any DES projects; therefore, it is essential for any successful DES project to develop process models that are reusable and close reflection of reality. Using process mining mitigates these issues by automatically discover process models from event logs [17]. Through the application of process mining techniques, healthcare organizations can discover the actual patient pathways that are conducted in reality [14], understand the high variance in clinical pathways taken by diverse groups of patients, and gain insights into bottlenecks and resource utilization [13].

As process models form the foundation for a simulation model and based on the abovementioned merits, this paper demonstrates how process mining can be applied to support key modeling efforts by the discovery of process models from historical event. Process mining techniques are used to analyze the event log from an Emergency Department (ED), identify the critical resources and performance bottleneck.

2 Methodology

2.1 Project Background

The hospital studied in this paper is an acute, public, voluntary, and adult teaching hospital that holds a unique place in the delivery of healthcare not only to the community of North Dublin but also to the rest of Ireland. This 570-bed hospital provides primary, specialized, and tertiary healthcare services, with a 24 h ‘‘on-call’’ ED which services over 55,000 patients annually. According to the task force report in 2007, the overall ED physical space and infrastructure is inadequate. Additionally, the hospital is operating at approximately 99% occupancy with resultant difficulty in accommodating surges in numbers of ED admissions. This is often aggravated by delays in patients transfer to critical care (ICU/HDC) beds. Consequently, the hospital is not compliant with volume and wait time targets (6-h patient experience time target). A detailed simulation model for the ED was developed in [1, 2] and a number of improvement strategies were proposed to achieve this target. Although these strategies were effective, the model was not flexible to accommodate the constant changes in patient care pathways and to sustain improvement efforts. In order to overcome these issues, the process model of the ED (that was developed manually) has to be updated to capture the changes in patient flow. Following the manual process of developing and updating the ED process model would take 6–8 weeks [1, 2]. Given the fast changes in healthcare process, by the time the model is completed the process model will not be reflective. Therefore, in order to minimize the latency between the occurrence of events and decision making, process mining techniques were applied to automatically discover patients’ pathways from historical data. A collaborative project with the hospital was established in order to achieve the following objectives: (1) discover the actual patient pathways that are conducted in reality; (2) understand the high variance in patient pathways taken by diverse groups of patients; (3) gain insights into bottlenecks and resource utilization; and (4) minimize the latency in the decision making process.

2.2 Dataset and Methods

A real-time patient tracking information system was used to track the patient’s journey within the ED. A 1-year historical data with anonymous patients’ records have been provided by the hospital managers. The dataset was provided in an event log structure with a total of 229,971 event logs representing 40,777 patients. Each log in the table represents an event (i.e. one process stage of the patient journey in the ED) with the following attributes (patient ID, Triage Category, Presenting Complaint, Date of Birth, Gender, Event ID, Tracking Step Name, Tracking Step Date Time, Location, By (Staff)). Events with the same name, patient ID and timestamp were removed. This resulted in 210,180 event logs.

Due to the unstructured nature of ED processes, the fuzzy miner [10] is used in this paper to discover the ED patient flow and to assess the variability of the flows within the process. The fuzzy miner allows to observe complex processes at different levels of granularity. This is achieved by applying two fundamental metrics: significance and correlation [10]. The significance metric assesses the relative importance of a precedence relation between two event classes, i.e. the more often two event classes are observed after one another, the more significant their precedence relation. On the other hand, the correlation metrics indicates how closely two events (i.e. activities) are following each other. Therefore, fuzzy mining could reduce and focus the displayed event classes by applying the two metrics on the discovered process map to achieve different levels of aggregation and abstraction.

3 Patient Pathway Discovery and Analysis

The main flow unit for the ED is the patient. A flow unit can be defined as an entity that enter the system, where various activities are performed before exiting the system. The event log for patients was analyzed to extract patients’ characteristics and types. Upon their arrival, patients are assigned a clinical priority (triage category) according to the Manchester Triage System (MTS) that is widely used in UK, Europe, and Australia [6]. The MTS uses a five level scale for classifying patients according to their care requirements; immediate, very urgent, urgent, standard, and non-urgent. Patients were grouped based in their triage category. Immediate and very urgent patients represent 15%, urgent patients (triage category 3) represent the largest group of attendees to the ED annually (59% average), while standard and non-urgent patient 26% of all patients. As advised by ED consultants, the analysis of these patients’ groups are critical as each group of patients can have a different journey within the ED and hence a different pathway.

3.1 Patient Pathway Discovery

The main building block of patient pathways are the activities that patients go through their journey in the ED. Twenty-two different activities within the ED were identified from the event log data. The fuzzy miner was then applied on the whole event log to construct the first top-level process map of the overall ED (Fig. 1a). The resulted map is too complex and is not interpretable due to the high variances in patients’ pathways.

Fig. 1
figure 1

The discovered patient flow model of the emergency department

Full size image

This confirms the perception of the complex nature of patient journeys within the ED, there will always be patients presenting to the ED with a unique characteristic that would require the patient to follow a unique care pathway. Furthermore, this complexity is what doctors and nurses deal with on a daily basis and it is one of the main reasons they do not believe that system engineering techniques can contribute to understand the complexity of patient flow. However, the fuzzy miner allows to observe complex processes at different levels of granularity. This is achieved by applying two fundamental metrics: significance and correlation. The significance metric assesses the relative importance of a precedence relation between two event classes, i.e. the more often two event classes are observed after one another, the more significant their precedence relation. On the other hand, the correlation metrics indicates how closely two events (i.e. activities) are following each other. Therefore, fuzzy mining could reduce and focus the displayed event classes by applying the two metrics on the discovered process map to achieve different levels of aggregation and abstraction. The fuzzy process miner was therefore applied on the data to show the main highway paths for patients and to hide less frequent paths (Fig. 1b). The number inside the rectangle shows how many times an activity has been executed. For instance, activity ‘Doctor Seen’ occurred 31,571 times. The number on the arc represents the co-occurrence frequency between any two activities. For example, the co-occurrence frequency between ‘Doctor Seen’ and ‘Referred for Admission’ is 7,205. Due to excluding low frequent paths there are differences between the numbers of activities shown on incoming arrows and activity boxes on the process maps. Further analysis of this model revealed that there were 1,984 different patient pathway patterns (Table 1).

Table 1 Discovered patient pathway patterns
Full size table

Over 60% of these patterns are one-off paths and only 31 of these patterns account for 80% of ED patients. Therefore, the remaining 1,951 patterns, which accounts to 20% of patients, were filtered out in order to reflect the common behavior of the ED. The fuzzy miner was applied again on the resulted 31,447 patients to drive final top-level process map of the ED (Fig. 2). The most followed paths are shown with thick arcs between activities. However, the analysis of exceptional pathways (paths with very low frequency) can give deep insights for medical professionals regarding the main factors behind these patterns.

Fig. 2
figure 2

The top-level process map of patient pathways

Full size image

3.2 Performance and Bottleneck Analysis

By considering the timestamp of events in the dataset, the ED performance and bottlenecks can be identified and analyzed (Fig. 3). The numbers on the arcs represents the waiting time between any two activities. The average length of stay (LOS) for all patients from arrival to departure (whether discharged or admitted to the hospital) is 9.1 h which is 3 h above the national target of 6-h average length of stay (LOS) in Ireland. However, the waiting time for admitted patients is 14 h on average. Patients have to wait 3.3 h on average to be seen by a doctor and 5.1 h afterwards to be discharged from the department. The main bottlenecks in the ED are the “Seen by Doctor” and “Patients waiting admission” activities.

Fig. 3
figure 3

Performance analysis of the top-level process map

Full size image

To gain a deeper understanding of the process flow of patients and the causes of these bottlenecks, the process model was analyzed at a more fine-grained level. The “Triage Category” attribute was used to divide patients into three groups; Immediate and very urgent, non-urgent and standard, and urgent patients. The process map of each patient group was constructed using the fuzzy miner and pathway patterns that reflect the common behavior for each group was analyzed (Fig. 4). There are obvious variances in the associated pathways for patients with different urgency categorization (i.e. triage). The first patient groups (Immediate and Very Urgent) represents 15% of all patients with the majority of them have been admitted to the hospital with an average waiting time of 13.7 h, for the admission process to be completed (Fig. 4a). While 26% of all patients are Standard and Non Urgent patients whom have a shorter pathway with a discharge outcome and 5.1 h average length of stay. The “Did not Answer” activity for this group represents patients who left the ED after being triaged without waiting to be investigated by a physician (whether a doctor or advanced nurse practitioner) (Fig. 4b).

Fig. 4
figure 4

Performance analysis for patients with different triage categories

Full size image

Urgent patients represent almost 60% of all patients with 10.1 h average LOS. This group are presented to the ED with a wide range of complaints with 27% are referred for admission and the remaining are discharged with an average waiting time of 5.2 h (Fig. 4c). The insights from this analysis enabled the ED decision makers to identify the bottlenecks for each group of patients and the challenges that they need to address.

3.3 Staff and Resources Analysis

Two types of resources were provided for each event; location and staff type. Therefor resource requirement was analyzed for different activities in patients’ pathways (Table 2).

Table 2 Resource analysis for the main activities in patient pathways
Full size table

The resource analysis gives deep insights regarding the gap between the guidelines that should be followed and what is actually happening. For example, the triage activity should take place in the triage room by a registered nurse (RGN). However, the analysis reveals that 68% this activity takes place in the triage room and 77% of the times is performed by the RGN. This is a clear evidence of the overcrowding of the ED and also quantify how fare this activity from the guidelines. Similarly, the “Doctor seen” activity is performed by senior hospital officer (SHO) (58%) and registrars (20%) in the Majors area in the ED (55%) and in the Resuscitation room (18%). This highlights the actual time spent by doctors in this activity and the actual locations where this is happening. These insights helped the ED managers to understand the actual allocation of staff and resources within the ED and to identify the gaps between best practices and the actual performance.

4 Discussion and Conclusion

The different views of the analysis provided in previous sections (patient pathways, performance and bottlenecks, and resources) enabled the ED managers in the hospital partner to discover the actual patient pathways that are conducted in reality, understand the high variance in patient pathways taken by diverse groups of patients, and gain insights into bottlenecks and resource allocation and utilization. In general, healthcare processes are highly complex, dynamic, and ad hoc in nature. Therefore, there is a need for techniques that can cope with the intricate complexity of healthcare systems. Healthcare information systems generates event log data that track patient-care processes as they take place are a valuable source of data for analyzing and studying these processes. Based on these event logs, process mining techniques were used in this paper to discover patients’ pathway patterns that are consistent with the observed dynamic behavior. Therefore, a more accurate process model for patient journey were developed and hence more adequate decision support. A current project with the partner hospital is in progress in order to integrate the process mining engine with the Hospital Information System to provide a real-time tracking of the ED processes and to minimize the latency in the decision-making process.