3.1 Research strategy

A mixed-method approach was adopted, involving: (i) an exploratory survey for the assessment of the complexity guidelines in ICUs, addressing the micro and meso levels; (ii) semi-structured interviews with health care professionals representing the micro, meso, and macro levels; and (iii) non-participant observations of the regular meetings of the ICU bed management committee in a capital city in Southern Brazil.1 These observations contributed to the assessment at the macro level. The joint analysis from all data sources supported the identification of five lessons learned, one for each complexity guideline. Data collection occurred from late September 2020 to mid-February 2021, which encompasses the tail of the first wave and the second wave in Brazil but not the explosive rise in cases during March and April 2021. There was no formal implementation of the guidelines in the settings where this study was undertaken. Thus, our research might be framed as an assessment of the studied settings against the domains identified in the guidelines, which are thought to be supportive of resilient performance to cope with complexity. The research project was approved by the ethics committee of the teaching hospital linked to the first author's university and informed consent was obtained from all participants.

3.2 Survey questionnaire

The questionnaire was made available on the Survey Monkey platform and a pilot application was conducted with three ICU professionals who were personal contacts of the authors. Next, it was sent out to potential respondents through three main channels: (i) participants of the previously mentioned ORCHESTRA project—this invitation was made by the former leading researcher of that project; (ii) a list of 25 ICU chief-physicians in the state of Rio Grande do Sul, as one of the authors was a former coordinator of the regional chapter of the society of intensive care physicians; and (iii) personal contacts of the authors. Despite these efforts, there were only 33 valid responses2 (all from different ICUs) from 13 out of the 27 Brazilian States. Three responses were discarded as they were provided either by professionals with too little ICU experience (<2 years) or the respondent clearly did not understand the use of the scale, assigning scores of only zero or 100 to all questions. Respondents had on average 14.6 years of ICU work experience, ranging from 3 to 38 years. Their distribution according to professional group involved: 23 physicians, 8 nurses, 1 physiotherapist, and 1 nutritionist—16 respondents had a team or unit management position.

The sample size was too small to produce statistically generalizable results. As such, the purposes of the survey were: (i) to provide an exploratory assessment of the ICU characteristics against the guidelines; (ii) to work as a script for the semi-structured interviews (see Section 3.3); and (iii) to indicate salient issues (e.g., survey statements with high or low scores) that could be explored in more detail jointly with the other data sources, creating a chain of evidence. Thus, sophisticated statistical analysis was neither possible nor essential for answering our research questions. A similar approach, which did not give prominence to statistical analysis, was used by Righi and Saurin (2015) when using a questionnaire for assessing the guidelines in an emergency department. On the basis of these premises, the survey results were analyzed using: descriptive statistics; Pearson's correlation coefficient between the groups of guidelines; and Cronbach's alpha for the 23 questions with the sliding bar. An alpha of .96 was obtained, which suggests a reliable instrument (Hair et al., 2014).

3.3 Interviews and observations

All survey respondents were invited to participate and seven agreed to be interviewed. Interview questions corresponded to the core part of the questionnaire (i.e., the questions on the complexity guidelines) with the aim of understanding the rationale for the survey responses. An additional joint interview was conducted with two public health officials who were also physicians and played a key role in the previously mentioned ICU bed management committee. Those interviews had two questions: Which are the roles and activities of the committee? Which are the benefits, difficulties, and criteria adopted for allocating patients to hospitals? The interviews lasted 1 h on average and all of them were audio-recorded and fully transcribed. Table 2 summarizes the profile of the interviewees.

Another source of data involved 20 h of non-participant observations (December 2020 to February 2021) in 10 virtual meetings of the municipal ICU bed management committee. Since April 2020, meetings have been held every week on Mondays, Wednesdays, and Fridays. The meeting involves 15 technicians from the municipal health department and focuses on the monitoring of a number of indicators, such as bed occupancy rates of ICUs, number of performed COVID diagnosis tests in the local health care services, and main outbreaks under way. Each meeting lasted on average 2 h and the researchers took notes in real-time when they detected comments related to the use of the complexity guidelines.

Interviews and non-participant observations were continued until data saturation criteria were met, which means that findings started being repetitive and the data produced was regarded by the researchers as sufficient for the purpose of answering the research questions (Ritchie et al., 2003).

The thematic analysis followed with the coding stage, in which excerpts of text were tagged according to the themes. Three researchers carried out a preliminary coding based on these criteria, separately. Then, codifications were compared in a joint meeting, inconsistencies were detected and a discussion followed until a consensus was obtained. This coding was still further reviewed by another researcher who read all transcripts and was a co-author of all of the earlier studies related to the five complexity guidelines, resulting in additional adjustments. The thematic analysis continued with the charting phase, which synthesized findings from the previous stages. For each complexity guideline, tables were developed (see Section 4), presenting the corresponding resilience practices. An overall analysis of the findings occurred at the mapping and interpretation stage.

4.1 Main characteristics of the ICUs

Table 3 presents the main characteristics of the 33 surveyed ICUs. They were mostly general ICUs (79%), small-sized (76% had no more than 20 beds), with interdisciplinary rounds (88%), organized as open-plan environments instead of single rooms (64%), part of teaching hospitals (64%), and with accredited quality and safety management systems (64%).

4.2 Provision of slack resources

Table 4 presents the results for the guideline on slack resources. The high incidence of the meso level (18 out of the 30 practices) reflects the fact that ICU slack resources were often borrowed from other hospital units.

The higher score (60.1) obtained by statement (5), which is related to the availability of material resources, might reflect the timing of the survey application, which occurred several months after the start of the pandemic and before the major rise in March/April 2021. All interviewees mentioned that the lack of supplies was more common during the early stages of the pandemic, when there were notorious problems of lack of PPE and drugs such as those necessary for the sedation of patients. Resilience practices that addressed these shortages included the replacement of the standard drugs by others with similar purpose and the development of new protocols for the sterilization and reuse of PPE, such as face masks.

In turn, the highest score related to this guideline (74.0) was obtained by statement (2), which was related to the dynamic management of material resources. The key dimension of this statement, across the micro, meso, and macro levels, was clearly related to the management of ICU capacity. The evolution of the pandemic, with highs and lows in the number of infected people, implied cycles of closing and opening ICU beds. At the micro and meso levels, common resilience practices involved the repurposing of facilities such as recovery rooms and the suspension of elective surgeries, which released not only material supplies but also staff. At the macro level, COVID hubs were created in some hospitals and patients with comorbidities (e.g., obesity) were directed to specialized hospitals. The two interviewed officials (#7 and #8) perceived that the public health system could effectively be managed as a network and make the best use of the overall capacity due to the centralized governance of the municipal bed management system, in contrast to the capacity management of private hospitals, which was fragmented.

Overall, the guideline on slack lies at the core of the pandemic response, which from the viewpoint of the provision of health care services might be framed as a misalignment between capacity and demand. Both capacity and demand changed frequently over the course of the pandemic, which posed the challenge of capacity being always ahead of demand, which in turn rose sharply at short notice. For instance, in 3 weeks during February/March 2021, the number of ICU beds occupied by COVID patients in the studied capital city rose from 282 to 866.3 A substantial portion of this addition of beds was only possible through the improvisation of non-ICU beds and the opening of ICU beds without full-time intensive care physicians.

4.3 Diversity of perspectives in decision-making

For similar reasons, interviewee #2 (chief-physician) remarked that, although there was a committee that canvassed the opinions of professionals, he would make the final decisions because he would be legally responsible for the outcomes.

This report acknowledges the importance of professional psychological support to staff − 12 out of the 33 ICUs had a dedicated psychologist—as well as to the role of leaders as moderators of the diverse perspectives.

Threats to the consideration of diverse perspectives were also identified from the qualitative data. One of these stemmed from the large number of inexperienced professionals hired to cope with the surge in demand. According to interviewee #3 (chief-nurse) “although nurse technicians attended a 2-week induction training, it was not uncommon that they were unable to perform simple procedures…furthermore, they were very insecure and frightened, asking for our help quite often; this took our time.” Interviewee #2 (nurse) made a similar remark: “many professionals were on their first job, no experience with critically ill patients, and of course no experience with COVID…then you can imagine, it was chaotic, terrible…experienced professionals (who worked in these areas full of new employees) did not want to return after seeing the chaos.” These reports indicate a downside of diverse perspectives; namely, when there is a wide knowledge gap between team members the benefits of the exchange of opinions and information is mostly unidirectional, from the most to the least competent people.

4.4 Visibility of processes and outcomes

Table 6 presents the results for the guideline on the visibility of processes and outcomes. The prominence of the micro level (12 out of the 14 practices) reflects the operational character of most visibility practices, which are directly targeted at the ICU staff. The two highest scoring statements of the whole questionnaire were related to this guideline. Statement (13) scored 89.2, which suggests that accessibility to information about the treatment and condition of each patient was not seriously compromised by the high number of patients. Similarly, workplaces remained clean and tidy in general as indicated by statement (11), which scored 80.5. These positive results might reflect practices that were solidly implemented before the pandemic (e.g., electronic medical records and housekeeping) and that resisted the scenario of scarce resources. In addition, 21 out of the 33 ICUs had accredited quality and safety management systems. Practices associated with housekeeping were probably in place partly due to accreditation requirements.

On the other hand, statement (12) had a fairly low score (51.0) suggesting that the results of performance indicators related to the overall ICU performance are not effectively disseminated. It is possible that this type of information, if available, is accessible mostly to managers. This is a drawback for resilience as the ability of the front-line professionals to anticipate threats and opportunities is hindered. Also, the rationale of decisions made by leaders may be unclear to other professionals as they do not have the whole picture of the ICU performance.

The role of digital technologies for the implementation of this guideline is also worth highlighting. In particular, despite the utility of electronic dashboards for displaying performance indicators, a drawback came to light during the observations of the meetings of the municipal bed management committee. Amid contrasting views regarding the extent of the mobility restrictions to be applied, some committee participants who were in daily contact with the front-line of care felt that some of the top health officials did not fully grasp the extent of the dramatic situation experienced at the front—one of the meeting participants reported that it was necessary “to see with their own eyes” in addition to looking at the dashboards that guided the meetings.

4.5 Monitoring and understanding work-as-done

Table 7 presents the results for the guideline on work-as-done. There were 11 practices or conditions, all of them related to the micro level. This is expected as the work-as-done of interest to this study occurred within the ICU premises. Statement (17) obtained the highest score (67.1), which may reflect the use of voluntary incident reporting systems as a relatively common practice in health care services.

By contrast, statement (18), concerned with learning from what goes well and normal everyday variability, had the second lowest score (44.5) of the whole questionnaire. This is unsurprising as health care services are known for their reactive safety management approach, which focuses on learning from failure (Braithwaite et al., 2020). Furthermore, this drawback is possibly part of a broader lack of emphasis on understanding work-as-done as indicated by the low score of statement (19)—46.2, which is concerned with studying work-as-done before making changes in management and care protocols.

Interviewee #9 (physiotherapist) provided another learning example related to prone positioning: according to her, the time taken to prone a patient reduced from 40 to 15 min during the pandemic.

In turn, the low score obtained by statement (15)—52.7, suggests that the gap between work-as-imagined and work-as-done may have grown wider during the pandemic. This is hypothesized to have occurred for reasons such as:

4.6 Monitoring and understanding unintended consequences

Table 8 presents the results for the guideline on unintended consequences. The survey scores were in general low and the overall mean was the lowest (45.4) among the five guidelines. To some extent, these low scores reflect the difficulties of the pandemic period, in which overloaded professionals had little time for the analysis of barriers and risks of changes (statement 21, score 46.1) as well as for collecting data on performance indicators (statements 22, score 46.8 and 23, score 44.5). In turn, the low score of statement (20)—44.1—is particularly worrying as rapid cycles of small-scale experimentation would be even more relevant in face of changing pandemic demands.

Unintended consequences were frequently a result of using the guideline “provide slack resources.” This makes sense as slack changes the nature of the interactions between the system elements (Perrow, 1984). For instance, staggered times for using the ICU staff room prevented gatherings (i.e., it created slack) but hindered informal social interactions between caregivers, which could be useful for resilience.

Slack also implied unintended consequences when resources were transferred from one part of the system to another. This is exemplified by the suspension of elective surgeries and outpatient consultations during the most critical moments of the pandemic; this freed up staff and facilities to treat COVID patients. Interviewees #7 and #8 (public health officials) stressed that the health condition of many non-COVID patients deteriorated due to the postponement of clinical care, which led to a demand surge after the regular procedures were reinstated, adding to the high workload from COVID patients. Overall, it seems that the observed unintended consequences were not necessarily unexpected but accepted as a price to be paid for short-term goals. Decision-makers either implicitly or explicitly opted for the course of action judged as most effective and morally acceptable in face of the circumstances.

5.1 What resilience looks like in ICUs during the pandemic

This study shed light on what resilience looks like in ICUs during the pandemic. Although 70 resilience practices were identified, several of these are not applicable to everyday complexity (e.g., addition of several new beds at short notice). This point reinforces the need for understanding what resilience looks like in a pandemic as the corresponding adaptive strategies differ when compared to other types of disasters—for example, contagion by infectious diseases may not be a serious concern when coping with victims from mass casualty accidents.

Nine unintended consequences surfaced, meaning that the resilience practices were at best approximate adjustments (Hollnagel, 2012) due to the high interconnectivity and trade-offs that characterize health services (Braithwaite et al., 2020). Figure 1 presents the total number of practices and unintended consequences according to the micro/meso/macro levels and the corresponding guidelines. Although the data in Figure 1 is not generalizable, it illustrates the potential utility of this type of analysis in larger data sets, provided any patterns are identified from them—for example, slack resources are more likely to be found in other hospital units (i.e., meso).

Some of the unintended consequences played out at the macro level as a result of practices deployed at the micro and meso levels. Similarly, some practices at the macro level certainly influenced the two other levels. This makes clear that the performance of the three levels is intertwined and their isolated optimization is likely to be ineffective in the long term, which is consistent with the proposal of Berg et al. (2018).

However, despite all efforts, results (and knowledge publicly available on the pandemic) suggest that ICU resilience was not enough to support desirable performance and that running out of resilience was part of everyday work during a prolonged period. These features are not self-evident in earlier descriptions of resilience that mostly focus on either everyday work under “normal” times or short-lived crises (Patriarca et al., 2018).

The concept of running out of resilience was evidenced by: (i) the lack of effective treatments and the consequent high mortality rates compromised the production of the required outputs (i.e., healed patients), which is a key dimension of resilience (Hollnagel et al., 2013); (ii) professionals who could not cope with the stressful working conditions—Azoulay et al. (2020) reinforce this point as they found that the pandemic had an overwhelming psychological impact on intensivists, who experienced distress due to suboptimal decision-making and provision of substandard care; and (iii) mismatches between capacity and demand, which highlighted the finite nature of slack resources—this had also been noted by Anderson et al. (2016) as a key for resilient health care.

Thus, running out of resilience can be framed as instantiations that play out in the form of unwanted events such as an adverse event or a professional that resigns. Furthermore, resilience in disasters such as the pandemic tends to degrade slowly (Woods, 2015) as margins get smaller and smaller over the course of hours, days, and weeks, rather than following a step function. This proposition is supported by the model of critical care surge capacity planning proposed by Hick et al. (2014), which places pandemics at the extreme of operating conditions, requiring ICUs to expand at least three times (200%) their usual capacity for a prolonged period.

The reliance of health services on human performance, which is arguably more adaptive than technological artifacts, is likely to be a reason for the graceful degradation during the pandemic. A more general implication of this insight is that, similarly to what occurs with technical infrastructures (e.g., Dunn et al., 2018 on electricity networks, and Zobel & Khansa, 2012, on cyberinfrastructures), there might be patterns of resilience curves for different sectors and disasters types. The identification of these patterns might guide the prioritization of resilience practices (e.g., slow degradation offers more opportunities for learning while the events are still unfolding, while brittle step functions call for preplanned responses) as well as the identification of benchmarks for intra- and intersector learning.

5.2 Lessons learned in light of the complexity guidelines

The prominent role of the guideline on slack resources was clearly related to the mismatch between ICU capacity and demand. All sources of data indicated that qualified caregivers were the key slack resource as it was harder to obtain in comparison to other resources. ICU staffing shortages have also been a major problem in the US during the pandemic, in which contingency plans include the use of ICU telemedicine (Harris et al., 2020). Furthermore, slack resources were commonly found beyond the ICU boundaries such as staff from other hospital units and beds in other hospitals. In fact, the pandemic led to the creation of units and teams practicing intermediary assistance between ICU and non-ICU wards—for example, in non-ICU wards dedicated to the care of COVID patients, noninvasive ventilation devices and high-flow nasal catheters have been used in an unprecedented way (interviewee # 1). This situation corresponds to the concept of “ICU without walls,” which is based on two premises: (i) the collaboration of all staff involved in patient care during hospitalization; and (ii) technological support for the early detection of patients at risk of deterioration throughout the hospital, based on the assessment of vital signs and/or laboratory test values (Gordo & Abella, 2014).

Thus, the guideline on slack deals with a tangible dimension of complexity management that involves decision-making on the right amount and mix of resources (Spearman & Hopp, 2021). Additionally, slack can be either designed (i.e., a resource is planned ahead of time to cope with predefined variabilities) or opportunistic (i.e., a resource plays a role as slack even though that was not their original purpose) (Saurin & Werle, 2017). This backdrop sets the stage for the following lesson learned.

Lessons learned related to slack: Capacity addition to health care services such as ICUs must encompass a wide mix of designed and opportunistic slack resources, which are likely to be out of the boundaries of the service focused on. Caregivers are likely to be the major slack resource as they enable the use of other resources and cannot be easily made available at short notice with the required competence level.

Regarding the guideline on diverse perspectives, it best represents the constructivist side of the adopted philosophical view of complexity. Both the advantages and disadvantages of diverse perspectives were probably amplified during the pandemic (Saurin, 2021) as the disease was new and there were many stakeholders (e.g., health care providers and businesses owners) with partly conflicting objectives. On the one hand, the use of diverse perspectives was exemplified by creative problem-solving and multidisciplinary care not only to patients but also to providers (e.g., counseling services and closer support from leaders). On the other hand, drawbacks of diverse perspectives were highlighted such as the creation of new meetings for the exchange of information and the effort spent to fight the spread of misinformation originated from the external environment (e.g., unproven treatments)—these drawbacks are costs of collaborative work (Goorden et al., 2014).

The analysis of this guideline can also be conducted in light of the law of requisite variety that applies to complex systems. This law states that a system can only be stable if the number of states of its control mechanisms is equal to or greater than the number of possible states of the system (Ashby, 1991). The possible system states, in this case, might be represented by the profile of the patients, which were cared for in ICUs dedicated to a single disease (i.e., COVID). It is hypothesized that this relatively low variety of patient profiles demanded a disproportional variety in terms of medical viewpoints, tentative treatments, and support from other health care services. Thus, it seems that the law of requisite variety misses the novelty of the system state—that is, if variety is low, but novelty is high, the control mechanisms should be both varied and novel. The lessons learned on this guideline are presented below.

Lessons learned related to diversity of perspectives: the benefits of diverse perspectives apply not only to patient care but also to the well-being and safety of providers. Furthermore, novel situations make the diversity of perspectives even more important, despite the possible low diversity of the process to be controlled. Also, a portion of the coordination costs stemming from this guideline involves the resources spent to fight against perspectives that add unnecessary complexity.

In turn, the guideline on visibility took advantage of both digital technologies (e.g., WhatsApp groups) and work organization measures (e.g., housekeeping). However, the role of the former approach was not strongly emphasized by the interviewees, which is to some extent contrasting with the potential of digital technologies for supporting resilience (Borsci et al., 2018; Tortorella et al., 2021).

The statements related to the visibility guideline obtained the highest average overall score in the survey, which may be due to reasons such as: (i) the importance of visibility under conditions of high uncertainty (Beynon-Davies & Lederman, 2017); (ii) the invisible nature of the main hazard (i.e., the virus) which makes the visibility of proxy indicators (e.g., patient with symptoms, number of COVID tests) even more important than in normal times (Saurin, 2021); (iii) the pre-existence of solid visibility practices as a result of past improvement initiatives; and (iv) the large number of inexperienced employees, who can benefit even more from easy access to reliable information. As a drawback, visibility can be a source of additional stress (Bernstein, 2017) to caregivers as they are aware of the deteriorating condition of patients and systems and can do little about it. On the basis of this context, the lessons learned on visibility are summarized below.

Lessons learned related to visibility: Reliable, real-time, accessible, and easily interpretable information on the status of health care services are realistic goals during crises such as the pandemic. Digital technologies and quality and safety management systems help make those goals achievable. However, a large number of inexperienced professionals demand visual management strategies tailored to their needs. Furthermore, the availability of information must be accompanied by action-taking; otherwise visibility practices can be discredited and become a source of frustration.

As for the guideline on work-as-done, the novelty, resource scarcity, and time pressure posed by the pandemic were catalyzed for learning by doing (e.g., prone positioning and intubation timing). Similar situations occur in other sectors such as in aviation (Carim et al., 2016), even though the corresponding crises are acute rather than chronic. As a result, work-as-imagined was probably frozen in its prepandemic form while work-as-done evolved at a fast pace. This learning experience is likely to inform the design of more realistic standardized operating procedures and training programs post-pandemic. However, the high workload and insufficient staffing were hindrances for the full exploration of the learning possibilities. Indeed, slack of time is critical for reflection and learning (Lawson, 2001)—statement (3) was related to the said slack and had a low score (58.2). The lessons learned related to this guideline are summarized below.

Lessons learned related to work-as-done: learning by doing in face of resource scarcity and novelty such as in the pandemic is to some extent inevitable and desirable. However, monitoring and learning from work-as-done might be even more important during crises, although that may require slack resources such as external consultants, staff from units not directly involved in the crisis, and spare time for reflection.

The guideline on unintended consequences shared commonalities with the guideline on work-as-done in terms of the low survey scores and implementation difficulties due to effort required for data collection and analysis. Despite these difficulties, major unintended consequences at the macro level stemming from the use of the other guidelines were certainly anticipated by policy-makers such as the creation of a backlog of patients with untreated diseases due to suspended elective procedures. In principle, awareness of these consequences would make it possible to mitigate their impacts when they played out. On the other hand, there were more subtle unintended consequences at the micro level—for example, lack of informal social interactions as a result of limits to gatherings in areas, such as staff rooms; inadequacies in the built environment of areas where new ICUs were installed. The corresponding lessons learned are summarized below.

Lessons learned related to unintended consequences: similar to the mitigation strategies deployed at the societal level (e.g., lockdowns), ICUs and hospitals also adopted simple measures in the sense of stopping activities such as elective surgeries, which drastically and quickly reduced complexity. Therefore, unintended consequences had a fractal nature (Song et al., 2006) as their nature was similar across scales—for example, hospitals lost revenue due to canceled surgeries similarly to businesses losing revenues due to lockdowns; caregivers suffered pandemic fatigue while there was an expected growth of mental illnesses in the population in general (Ornell et al., 2020). Public acknowledgment of this fractal nature could foster empathy in society and tackle the counterproductive perspectives of some agents. An active search for nonobvious unintended consequences stands out as another learning opportunity. That search benefits from the imagination of the work system designers and risk analysts (Adamski & Westrum, 2003), particularly in all decisions that involve the provision of slack resources. Slack has a disruptive potential by adding new elements (e.g., beds), reallocating elements (e.g., staff), or separating existing elements (e.g., social distancing). As a consequence, slack amplifies the potential for unintended consequences (Perrow, 1984).

Finally, as an overall lesson learned, the joint use of the guidelines seems to be desirable to resilient performance. This lesson is supported by Pearson's correlations between the guidelines, which were all positive and statistically significant at the 99% confidence level (Table 9). Righi and Saurin (2015) found similar results regarding these correlations. Moreover, these correlations make logical sense and were reinforced by data from the interviews and observations. For example, resilience practices related to visibility (e.g., whiteboards displaying information on the patient condition) can contribute to the free sharing of information across professional groups, thus facilitating decision-making based on multiple perspectives. In the same vein, resilience practices related to the monitoring of work-as-done (e.g., realistic simulation and training) can make caregivers more aware of the ICU dynamics, helping the anticipation and monitoring of unintended consequences of improvements and changes.