Do healthcare services behave as complex systems? Analysis of patterns of attendance and implications for service delivery

We analysed primary data from two sources: PCOOH data from a study of NHS 24, the service which provides out-of-hours primary care services throughout Scotland (population 5.6 million (M)) [35] and ED data supplied by Canterbury District Health Board in New Zealand. The data was for the ED of Christchurch Hospital, serving a population of approximately 500,000 people. Both datasets were derived from routine management data and thus included all cases handled by the respective services.

In the PCOOH service, all calls were initially managed through a nurse-based triage system with a range of options including telephone advice by the nurse, consultation with a general practitioner (GP), either at a treatment centre or in the patient’s home, and direct ambulance transfer to an ED. The data included all calls to the NHS 24 service throughout 2011. We excluded calls during office hours (08.00 to 18.00 weekdays except public holidays) because the vast majority of urgent care requests during these hours go directly to the patient’s GP practice. All data was anonymised and handled under a data-sharing agreement between the University of Aberdeen and NHS 24; as no patient-identifiable data was involved, additional research ethical permission was not required. The ED data comprised anonymised data of adult attendance for the period July 2011 to June 2013. Since no identifying patient information was involved, research ethical permission was not required under New Zealand’s Standard Operating Procedures for Health and Disability Ethics Committees.

For both datasets, the unit of analysis was a contact — defined as one or more attendances or phone calls with the service on a given day. While some patients had more than one attendance or call on the same day, we limited data to one contact per day, since most repeated calls on the same day related to the same episode of care (e.g. escalation of a contact from telephone advice to consultation). PCOOH contacts included episodes of care which were managed by telephone, in the urgent care department or at home, and by any healthcare professionals. ED contacts included emergency attendances by any route.

Heavy tails are a feature of a number of statistical distributions, including the inverse power law and the log-normal, which have been repeatedly observed in large, naturally occurring systems [11]. The term refers to the long, or thick, “tail” of the distribution, in contrast to the shorter tails of distributions such as the normal, where data points cluster close to the mean (and almost always within a few standard deviations of it). One of the first described heavy-tailed distributions was the Pareto principle whereby 80% of the wealth is held by 20% of the population (and of that wealth, 80% is held by only 20% of the 20%, and so on) [16]. This observation is mathematically related to the inverse power law. The result of such a distribution is that the range of wealth from poorest to richest is much greater than would occur if individuals were normally distributed around a mean. While plots of the normal distribution produce the familiar bell curve, conventional plots of heavy-tailed distributions are difficult to interpret because of the much larger range of values. A simple solution to this is to plot cumulative distributions with logarithmic axes. Using this format an inverse power law will describe a straight line (from top left to bottom right). Heavy-tailed distributions are typically seen in theoretical complex systems and have been held to be a typical feature, or fingerprint, of complex systems [11].

While there are many possible explanations for the occurrence of heavy-tailed distributions [36], and their presence alone is not proof of the presence of a complex system, the more recent description of correlated bursts of activity within a complex system leading to heavy-tailed overall distributions [12] means that identifying both in the same data strengthens the evidence for the presence of a complex system. Correlated bursts of activity represent clusters of activity in time, with the distribution of burst lengths also following a heavy-tailed distribution [12]. This distribution of bursts means that long bursts arise more often than would occur by chance, and furthermore that bursts undergo a kindling process, whereby the longer a burst has gone on, the more likely it is to continue.

We conducted this analysis using the total number of contacts per patient in 1 year of data from each dataset. We plotted the complementary cumulative distribution function (CDF), defined as the proportion of patients whose total number of contacts was equal to or greater than each number of contacts between 1 and the largest recorded. This is preferable to plotting the probability distribution function, as the CDF is more robust against fluctuations in the tail of the distribution due to finite sample sizes [11]. Plots of the CDF were made with logarithmic axes, on which a power law distribution displays a straight line. The slope of this line is equal to one minus the scaling parameter of the power law distribution (for example, a scaling parameter of 3 produces a slope of − 2). We fitted inverse power law and log-normal functions to each dataset using maximum likelihood fitting with the poweRlaw package in R 3.3.2. We examined the closeness of fit of the data to a power law distribution, using the Kolmogorov-Smirnoff (KS) test. Where data was deemed different from a power law distribution (p value < 0.05) we examined whether the data was closer in fit to a power law or log-normal distribution using the Vuong test for non-nested distributions [37] at a significance level of 0.01. As naturally occurring distributions often include a power law only above a certain threshold [11], we repeated this analysis with a range of minimum values for the number of contacts per patient.

We examined bursts of contacts using the method developed by Karsai [12]. We defined a burst as a sequence of contacts in which the interval between each pair of consecutive contacts is less than a specified time window ∆t. We used a range of values for ∆t of 4 and 7 days for the PCOOH data and 4, 7 and 10 days for the ED data. For each burst we counted the number of contacts within the burst. We pooled the burst patterns across all individuals and limited the analysis to patients with between 4 and 52 contacts in order to exclude data with too few contacts to display bursts, or so many contacts that bursts were inevitable because the PCOOH data comprised only 1 year. We conducted the burst analysis on the PCOOH dataset (1 year) and on the whole ED dataset (2 years).

We systematically identified papers reporting numbers of patients attending EDs, using a structured search of the MEDLINE and Embase databases (Table 2 lists the search terms). Titles, abstracts and full manuscripts were screened by two authors independently to identify usable data. We followed references from, and citations of, eligible papers to identify additional studies.

Primary data was available from 724,921 PCOOH patients (1,085,796 contacts) and 60,106 ED patients (98,228 contacts). The age and sex characteristics and number of contacts per patient are listed in Table 3.

Plots of total number of contacts per individual are shown in Fig. 1a (PCOOH data) and 1b(ED data). Both plots show a heavy-tailed distribution, which approximates to an inverse power law (straight line) for the whole distribution in the ED data and from approximately 5 contacts to 30 contacts in the PCOOH data. Above 30 contacts in the PCOOH data (Fig. 1a) the tail of the distribution can be seen to deviate from the power law; there were more patients than expected with very high numbers of contacts: 225 patients (0.03%) had more than 30 contacts. This represents approximately twice as many as would have been expected if the data followed a power law distribution. This pattern is suggestive of more than one overlapping distribution. Figure 1c shows the result of the sensitivity analysis in which the PCOOH data was split into patients whose first contact occurred within the first 14 days of the year and those whose first contact came later. The rationale was that patients consulting in the first 14 days might be within a burst of consultations at the start of the data collection and thus might be more likely to have repeat consultations than those starting their first burst after at least 14 days of no contact. The two resulting distributions in Fig. 1c both showed close approximation to a power law. Finally, Fig. 1d shows the analysis repeated with censoring of data at 6 months after the first consultation, indicating that this had no adverse influence on the observed distribution’s approximation to a power law.

Plots of the distribution of burst length are shown in Fig. 3a and b for both the PCOOH data and ED data. Both plots use a 7-day window for inclusion of contacts within bursts. Both distributions are clearly heavy tailed, approximating to a straight line indicative of an inverse power law. None of the 250 surrogate datasets, in which the temporal structure of bursts was disrupted, showed this distribution, suggesting that it was unlikely to have arisen in the data by chance. Similar patterns were seen from the PCOOH data with a 4-day window (Fig. 3c) and from the ED data with 4- and 10-day windows (Fig. 3d). This similarity across different time windows makes it unlikely that the observed results were due to an artefact of the measurement parameters and more likely that these new findings represent real phenomena present in the data.

This study used large, recent and complete datasets from two different urgent care settings in different healthcare systems. The analysis used established techniques for burst estimation [12] and model fitting [11]. We also adjusted for different lengths of follow-up by censoring data and found it had no influence on the findings. Examining the combination of both burst analysis and overall distribution analysis is important, as bursts have been identified as a generative mechanism for power laws in other systems. Furthermore, bursts have been identified in other healthcare research, such as the tendency of exacerbations of chronic obstructive pulmonary disease to cluster in time [52].

While the ED data showed a close fit to a power law across the whole range of contact numbers, there was some evidence that the PCOOH data contained more very frequent attenders (above 30 contacts) than expected from the best fitting model. This may indicate some excessive or inappropriate use, but the absolute number of patients was small. When we restricted the analysis to patients who did not use the service in the first 2 weeks of the year (and so who were not currently in a burst of consultations), the observed data showed a closer fit to an inverse power law.

The inclusion of the systematic identification of secondary data adds strength to our findings of overall distributions, as heavy-tailed distributions of use, similar to those seen in our primary data, were observed across very different healthcare settings, with generally similar parameters for the proportion of frequent consultation. We were not able to conduct statistical analysis on these secondary sources of data, as they did not have sufficient detail.

While complex systems have been hypothesised as a way of describing healthcare services [3‐5, 7, 13], this is the first large-scale empirical examination of whether urgent healthcare displays the typical statistical properties of a complex system. No previous studies have reported the population distribution of urgent care attendance in detail; however, non-normal distributions of use have been previously noted but not analysed in the ways we have used in this study [33, 34].

To be plausible, our finding of the typical statistical properties of a complex system must be compatible with real-life mechanisms, which in modelling of social systems can be considered as rules [5, 53]. Qualitative studies have already suggested candidate rules: patients simultaneously seek to balance being a prudent user of services [29] with being “better safe than sorry” [30], and this balance is influenced by societal processes and norms [31, 32]. In turn, these rules may be mediated through processes such as candidacy (seeing oneself as an appropriate user of services) and recursivity (a tendency to repeat patterns of help-seeking which have been successful) [28]. Together, these processes — which are socially mediated — can be seen as comprising system-wide mechanisms which drive, and restrain, urgent care use by individuals.

Frequent attendance is commonly regarded as abnormal and taken to be a sign of an inefficient system, however many frequent attenders appear to use healthcare appropriately [17], suggesting the system may in fact be operating efficiently. Recent work in information theory suggests that power law distributions may represent an optimal configuration for a system to meet very variable demands [54]: in the case of urgent care, systems must deal with many patients with minor problems while retaining the capacity to handle a few with intensive ones. Heavy-tailed distributions of attendance may be a feature of well-optimised urgent care rather than a sign that something is wrong.

Our findings of striking similarity between data from urgent care use and statistical features of typical complex systems support the argument that services need to engage more with a complex systems approach [3]. This means there should be a greater focus on contextual matters across the whole system and a recognition that the mechanisms driving processes such as demand both arise from, and influence, many individual interactions. In turn, this means there is a need for interventions to influence these mechanisms, which are social, both through information channels and media, and through creating and sharing positive patient experiences. A second general consequence of considering healthcare systems as complex is that interventions to change services must recognise that complex systems respond unpredictably to interventions to change them [3, 5, 15], and that what works in one setting will not necessarily work in another. This dependence on context is still under-acknowledged in the development of “complex interventions” [3], which should be viewed as “interventions in complex systems” [55].

In practice, the implication of our findings for frontline care is that there must be a partial shift in thinking from individual frequently attending patients to the workings of the whole system. While each frequently attending individual is unique, the consistent and mathematically predictable frequency with which they occur is highly suggestive of overall system effects. In theoretical models of complex systems, this dependence on system effects means that even if extreme outliers (such as individual frequent attenders) are removed (representing action on individuals), new ones will arise to fill their place [56]. This phenomenon can be seen in waiting lists — whereby initiatives to shorten them (by bringing forward treatment of individuals) generally lead to them rapidly re-growing through system effects [57, 58]. Services thus need to provide care which is simultaneously person-centred and system-aware.

For research, our identification of bursts represents a potential target for interventions to identify and respond to individuals with high need. Interventions should be developed to prevent, or shorten, bursts. These interventions must be safe, while addressing the mechanisms by which patients rationalise decisions to consult, such as candidacy and recursivity [28]. This may involve forms of explanation or sign-posting which make patients more likely to use alternative management the next time a situation occurs rather than more likely to re-attend the urgent care service, as currently happens. A focus on recognising bursts at an early stage may also permit identification of individuals at high risk of frequent attendance. In our ED data, among people who attended at least four times in a year, a burst of 3 consultations each separated by no more than 7 days represented only 1% of bursts. In the PCOOH setting, bursts of 4 consultations each separated by no more than 7 days accounted for 1% of bursts. These may represent useful “early warnings” of emerging problems, and these and other potential signals of ongoing high use should be tested in further analyses.