Situated Action in Pre-Hospital Critical Care Dispatch: Identifying where and how Algorithmic Assistance might be useful in the daily work of specialist Emergency Medical Dispatchers

1.1. Motivation

• The study identifies context-specific design as important because AI performance at the bench does not ensure clinical benefit in situated workflows.It frames this gap as the AI chasm between AI promise and clinical benefit.Source: §1.1 · p. 2

1.2. Context; 1.3. Focus

• The study focuses on a two-person specialist team deciding whether to dispatch pre-hospital critical care, amid assessment, filtering, prioritisation, and competing demands on attention.The team monitors emergency calls and dispatches critical care teams when incident information indicates likely patient benefit.Source: §1.2 · p. 3, §1.3 · p. 3

1.4. Prior work

• Prior work describes the human-AI combination space and cases where impressive algorithmic performance did not improve real-world emergency-service decision-making.The paper positions situated task understanding and evaluation of the human-machine combination as relevant to design.Source: §1.4 · p. 4

1.5.1. Research Objective 1 - uncovering methods

• The first objective is to uncover staff methods by studying their work across clinical shifts, policies, procedural conflicts, and field experience.These accounts inform how dispatch decisions become consequential moments for the team.Source: §1.5.1 · p. 5

1.5.2. Research Objective 2 - identifying key decision points

• The second objective is to identify pivotal decisions surrounding dispatch, including their inputs, influences, and outflows.The inquiry focuses on antecedent decisions that enable or influence the dispatch decision.Source: §1.5.2 · p. 5

1.5.3. Research Objective 3 - locating where to assist

• The third objective is to locate where algorithmic assistance could reasonably work, building on understanding of staff methods and decision points.The paper proposes starting from context and involving situated stakeholders in consequential design deliberations.Source: §1.5.3 · p. 6

1.6. Contribution

• The paper combines immersion and structured observation to describe work-as-done, workflow steps, staff interactions, and the interdependencies among decision points.It uses this situated account to assess candidate locations for support and design requirements.Source: §1.6 · pp. 6–7
• The authors identify one candidate decision step as the most promising place to explore an algorithmic assist and argue for beginning with a specific situated problem.They frame contextual groundwork as increasing the chances of success and sustainability, and the prospect of patient benefit.Source: §1.6 · pp. 6–7

2.1. Immersion methods

• The study used iterative immersion to discover decision moments and staff methods together, because understanding either depends on learning the other.Immersion aimed to establish familiarity with team language, tasks, tools, and the influences that make actions intelligible within the team.Source: §2.1 · pp. 7–8
• Immersion combined observation over several shifts with meetings, training, documents, anonymised call lists, and discussions to understand staff methods in context.The researcher spent over 50 hours in the work environment, including five shift handovers, plus about 50 hours in discussions.Source: §2.1 · pp. 8–9

2.2. Observation setting

• Structured observation recorded 22 operational hours across three day shifts and two night shifts, after an initial five-hour session refined the recording codes.Observation positions respected screen privacy while allowing attention to staff interaction and, from one position, paper-note use.Source: §2.2 · p. 9
• Observation foregrounded speech and attentional patterns because staff navigate an updating call list while coordinating with colleagues and operational teams.The study treated these interaction patterns as the cognitive backcloth for screen-based work, without examining the specific screen interface.Source: §2.2 · pp. 9, 11, 12

2.3. Observation methods

• Coding evolved from interaction modes to forms of staff-to-staff interaction, then underwent axial refinement and checks for counterexamples and assumptions.The first hour used constant comparison; later coding refined solicited and unsolicited utterances and checked links across interaction types.Source: §2.3 · pp. 12, 14
• The construct-indicator review traced how intended measures become records and counts, checking whether coded interactions could estimate situational-awareness interactions.For this study, Ix denotes interactions concerning situational awareness at the call-stack, general-case, or resource level rather than a specified individual case.Source: §2.3 · pp. 12–14
• Staff interactions occurred about once every 2 minutes on average, while colleagues frequently anticipated one another’s focus and were ready to interact on the same issue.Attention rarely remained on a single task for long unless a call was reasonably close to a dispatch decision.Source: §2.3 · p. 15

3.1.1. Contextual grounding - the role of triage

• Pre-hospital critical-care dispatch selectively tasks specialist teams to deliver high-level interventions at the scene, unlike routine ambulance response to every call.Cases not selected for critical care remain within the wider ambulance service’s normal dispatch process.Source: §3.1.1 · p. 16

3.1.2. The critical care hub workflow

• At least three recurrent decision steps—two triage stages followed by dispatch—progressively reduce the calls considered at each stage.New information can prompt staff to revisit provisional determinations.Source: §3.1.2 · p. 17

3.1.3. The workflow in outline

• At step A, staff decide whether to open a call; the headline alone cannot justify tasking a field team, and event logs are rarely searched without a specific target.Opening reveals accumulated call information; further detail may be available in another view.Source: §3.1.3 · pp. 17–18
• At step B, staff may interrogate a case through live call audio, questions relayed to the call-taker, or follow-up contact with someone at the scene.Audio logs take too long for real-time review, so ongoing call audio is usually the primary source; staff may also call back after a call ends.Source: §3.1.3 · p. 18
• At step C, a clinician decides whether to dispatch; most dispatches follow interrogation, while rare immediate dispatches omit it for speed.The field team lead and pilot, with the pilot’s flying-condition decision taking priority, determine whether the team travels by air or road.Source: §3.1.3 · pp. 18–19

3.1.4. Dynamic sift and sort

• Triage combines classification by care pathway with prioritization, but provisional decisions and calls’ relative priority can change as conditions and the call list evolve.Staff may seek updates or proactively call the scene when a case could become suitable for critical care.Source: §3.1.4 · p. 19

3.1.5. Features of the call list

• Every decision sequence begins when a call appears on the list, while the list’s wider-service technology constrains change largely to a limited presentation-layer API.Because the process can be initiated many times every few minutes, list interactions can affect workflow and task efficiency.Source: §3.1.5 · p. 20
• The hub’s all-Wales list is longer and updates faster than regional lists; keeping one’s place becomes burdensome beyond 15 or 20 items and significantly interruptive above 50.Calls may move position when updated, adding distractor effects as staff switch attention to other screens or interruptions.Source: §3.1.5 · pp. 20–21

3.1.6. Challenges of the workflow

• A previous call-taker flagging feature produced many false positives and missed many cases, adding work without perceived benefit before its removal.Staff also continually switch between reviewing changing calls at headline level and focusing on an individual case, punctuated by dialogue.Source: §3.1.6 · pp. 21–22

3.1.7. Informational profile of the process

• At step A, standardized AMPDS codes can be ambiguous, while call-taker free text adds uncoded details but varies in amount and consistency.The selected protocol can shape which incident details are elicited and encoded; call-list free text uses abbreviated specialist vocabulary and may not follow clinical terminology.Source: §3.1.7 · p. 23
• Over half of calls are reviewed at step B with added call-taker, location, timing, and system data; dialogue with someone at the scene supplies the highest-detail information at step C.Shared awareness differs across colleagues and outside interlocutors, making the communicative context relevant to decision support.Source: §3.1.7 · pp. 23–24, §3.2 · p. 24
• Less than 1% of listed calls are eventually dispatched, and step C’s highest-detail information is available for less than 2% of calls.The process moves from thousands of low-detail listings toward a small number of highly detailed cases, while decisions remain subject to change.Source: §3.1.7 · pp. 23–24

3.2.1. Internal and external interaction

• Operational communication with field teams, hospitals, and specialists is the second most common group but accounts for less than a third of observable interactions.The other groups are interactions between the two hub staff.Source: §3.2.1 · pp. 25–26
• Situational-awareness updates dominate the interaction groups and provide a foundation for many case-specific decisions.They concern active or available resources, relationships among calls, resource conflicts, and call disambiguation rather than an individual incident’s details.Source: §3.2.1 · p. 26
• Specific-call exchanges are the third-largest group, while other internal hub interactions include joint-task coordination and frequent affirmations or negations.These short replies are counted in the other-hub group because placing them elsewhere would inflate those groups; they often confirm information already held, rather than fill an information gap.Source: §3.2.1 · p. 26

3.2.2. Final grouping of coded types

• The coding aggregates 27 interaction types into six groups, including non-task exchanges that sustain communication and rare operational-advice exchanges that share professional knowledge.Table 3 compares average hourly group counts across four shifts and shows consistent within-group distributions alongside day/night differences.Source: §3.2.2 · pp. 26–27, Table 3 · p. 27

3.2.3. Patterns detectable in the observation data

• During the observed night shifts, interactions were less frequent overall, while situational-awareness and operational communications formed a greater proportion of interactions than during the day.Together, those two groups occurred about every three minutes at night versus about every two minutes in day observations; the night shifts involved remote desk working and reduced call and flight activity.Source: §3.2.3 · p. 27
• Operational exchanges with crews and hospitals are often interrupted by events or degraded communications, and therefore use formalized acknowledgements, turn-taking, clipped grammar, and repeated phrases.The Trauma Desk is another important external operational contact; the hub covers that function overnight.Source: §3.2.3 · pp. 27–28

3.2.4. Unsolicited updates

• Hub staff provide many more updates to each other than they explicitly seek, especially for situational awareness; this anticipation is internal to the team.Across the interaction types, updates can convey clinical or logistical case information, resource status, timelines, or effects on related decisions.Source: §3.2.4 · pp. 28–29
• In the busiest observed hours on two shifts, staff differed in whether they explicitly sought situation-awareness updates or supplied them unsolicited, a contrast the authors relate to adaptive working styles.One staff member was shared between the shifts; the authors report that the pattern was also reflected over the whole shift.Source: §3.2.4 · p. 29, Table 5 · p. 29

3.2.5. Factors affecting staff interactions

• The observations show changes by time of day and day of week, but not their long-run evolution; system pressures and case mix also affect staff across the service.When specialist teams are held at scenes because no regular ambulance crew can relieve them, hub decisions increasingly have to protect specialist availability for specialist needs.Source: §3.2.5 · pp. 29–30, §3.2.6 · p. 30

3.2.6. Variation over time

• Morning handovers bring a hub-wide briefing on roles, available resources, overnight effects, consultant contact, and changes to patient routing as services come on shift.The ensuing activity may be call interrogation or field-team support, depending on team activation.Source: §3.2.6 · p. 30
• When all teams are available at base, staff can review the call list and interrogate calls; when teams are in the field, coordination can dominate while the list remains under continual monitoring.At busy times, the hub may support all four field teams across multiple incidents over several hours without a break.Source: §3.2.6 · pp. 30–31

3.2.7. Sources of variation

• Weather changes can shift aircraft use to road vehicles, altering travel times, transfer plans, and patient-transport logistics.Aircraft may land far from the scene, and road vehicles cannot convey patients; each external change has a counterpart in the hub’s work.Source: §3.2.7 · p. 31
• Remote working requires explicit communication and updates, while co-location permits non-verbal signals and shared viewing of information.Muted external calls and dropped audio can prompt conversational repair and leave staff needing to report updates after unmuting.Source: §3.2.7 · pp. 31–32

3.2.8. Shared information

• Unprompted utterances often reflect a shared task and changing informational context, while staff maintain that shared knowledge through ongoing updates and checks.Their attention to the call list and the details behind each item can differ, even when they are aware of some of the same information.Source: §3.2.8 · p. 32

3.2.9. Situated action - simplifying a complex process

• The workflow has three key steps: attending to a call on the list, viewing its details, and making a substantive dispatch decision.The information available changes across the steps, from the call-list line to recorded call details and then audio or, sometimes, video.Source: Figure 8 · p. 33, §3.2.9 · pp. 33–34
• Updates may reach staff outside the normal decision sequence, and a later clinical update can return a call initially marked for no hub action to consideration.A call can move from a lower priority, such as Green2, to a higher one, such as Amber1.Source: §3.2.9 · pp. 33–34
• The three steps distinguish clinical need from operational feasibility, while the core pathway separates recurring decisions from interruptions and ancillary events.The authors identify the steps, or their combination, as candidates for algorithmic support because information changes at each point.Source: §3.2.9 · pp. 34–35

3.3.1. Three candidate steps for decision support

• Users resisted ready-made dispatch suggestions, wanting informational overlap, source evidence, reasoning, and visibility into the intermediate steps.Step A information is relatively homogeneous and available in one electronic system; step C draws on heterogeneous, multimodal sources, and live data integration would be costly.Source: §3.3.1 · pp. 36–37

3.3.2. Interaction with an algorithm when time is tight?

• Algorithmic review could be intrusive when clinicians must verify what information was captured and where, especially if the system disrupts their view.Staff anticipated verbose, difficult-to-probe justifications rather than the concise, readily challenged exchange possible with a colleague; they considered later-step support unsuitable unless highly performant.Source: §3.3.2 · pp. 38–39

3.3.3. Limits to algorithmic situational awareness

• Dispatchers tailor decisions to crew capabilities and changing hospital or team capacity, but much of this real-time situational knowledge is held by people or on paper.Crew resources documented in systems can differ from those actually available, while capacity may change daily or hourly and would be burdensome to update in a machine.Source: §3.3.3 · p. 39

3.3.4. Algorithmic assistance in the workflow

• Clinicians doubted that an algorithm could interpret the whole context or ask clarifying questions in real time, favoring assistance in the early workflow instead.Background sounds and incident details can prompt targeted questions that help establish clinical meaning.Source: §3.3.4 · pp. 39–40

3.3.5. The learning problem - evaluating decisions in a dynamic context

• Time pressure makes review thresholds dynamic: competing calls change how many cases staff can view and how much detail they can examine.The value of reviewing a call depends on the estimated background list, making a fixed decision label or ranking unsuitable across contexts.Source: §3.3.5 · pp. 40–41
• Historic outcomes cannot provide objective ground truth: dispatch records miss false negatives and can conflate justified critical-care cover with inappropriate dispatch.Only fifteen of 2,000 calls were dispatched, and downstream outcomes and interventions may not reveal whether the original decision was sound.Source: §3.3.5 · pp. 41–42
• Step A is the strongest initial target: clinician feedback is conceivable there, time-consuming at B, and prohibitive at C.False positives from A also consume more substantial decision time at B, making B less suitable for initial experimentation.Source: §3.3.5 · p. 43

3.3.6. Support for situated decision-making

• Hub decisions occur amid parallel calls and coordination work, while the time available to view calls varies with workload and incoming demands.Staff may revisit changed calls, and paired working improves call-stack monitoring, but busy periods can force faster or incomplete review.Source: §3.3.6 · p. 44
• Assistance must fit frequent task-switching across four screens and multiple input devices while remaining quick to access, case-specific, easily interpreted, unobtrusive, and non-interruptive.Monitoring gaps can last from a few minutes to over half an hour, so support also needs to be quickly reviewable and amenable to efficient feedback, especially early on.Source: §3.3.6 · p. 45

3.3.7. Situated action - summary findings

• The workflow involves two triage steps to sift high call volumes before a third step makes the substantive dispatch decision, with trade-offs between filtering time and decision time.The specialist team seeks the less than one percent of emergency calls that can benefit from pre-hospital critical care.Source: §3.3.7 · p. 45
• Decisions have no fixed thresholds: evolving case details, monitored calls, and resource availability affect assessment and disposition throughout the sequence.Hub staff also monitor calls, prioritize and coordinate field teams, and provide critical care advice to ambulance crews and first responders.Source: §3.3.7 · pp. 45–46
• Unsolicited situation-awareness interactions dominate interaction types, helping staff maintain a shared mental model as cases and resource availability evolve.Their frequent, economical exchanges reflect mutual trust, but also indicate that situational awareness may decay rapidly.Source: §3.3.7 · p. 46
• Support for the substantive dispatch decision would require complex multimodal data and an equivalent of shared situational awareness; second-triage support must account for changing demands and resources.Suitable training data is also a challenge, while detailed call data is available for the second triage step.Source: §3.3.7 · p. 46
• Step A, the first triage, offers a ready-made call-list input and interface plus efficient expert feedback, although it has little case data and no conventional ground truth.Steps B and C have less clear opportunities for intervention in the workflow.Source: §3.3.7 · pp. 46–47

4. Discussion

• Immersion and observation provide a practical response to calls for ecological validity and earlier interdisciplinary engagement with situated workflows.The authors argue that this work helps researchers understand real-world context deeply enough to support productive interdisciplinary work.Source: §4 · p. 47
• Contextual inquiry can inform early design choices, interdisciplinary work, and near-live evaluation by clarifying where algorithmic contributions fit and what people do there.The authors connect this shared perspective to assessing whether a deployment will work sufficiently well and remain worthwhile in a real-world setting.Source: §4 · p. 47
• Ethnographic techniques can uncover opportunities and barriers to algorithmic assistance, while observation of situated practice reveals how workflow decisions interact.The researchers combined observation with informal interviews and discussions to uncover daily practices and details absent from management descriptions and documented procedures.Source: §4 · pp. 47–48
• Clinicians’ accounts of information and mental processes clarified how they relate field experience to call interrogations and documented headlines.These accounts enriched understanding of decision-making at different workflow steps.Source: §4 · p. 48
• Clinicians’ strong preference for assistance at the first triage step led the authors to focus subsequent experimental exploration exclusively on that part of the workflow.This choice followed initial concern about organisational friction from back-end data that was not readily exposed through an application interface.Source: §4 · p. 48

5. Conclusions

• The study’s contribution is observational fieldwork demonstrating that designing for situated decision-making requires granular understanding of decision-makers’ situated action.The authors describe attentive observation and sufficient time embedded in staff practices as essential, even when ethnography is neither complex nor extensive.Source: §5 · p. 48
• The proposed intervention is an unobtrusive simple score in the existing call list, supporting the early decision to view within a multi-step triage process.The authors say the observational work informs both where support is provided and the form it takes.Source: §5 · p. 49
• The authors contend that situated-action approaches can apply beyond this clinical context because their conclusions focus primarily on sociotechnical rather than domain-specific elements.They argue that studying situated action helps determine whether techniques are practically valuable in another context or need modification.Source: §5 · p. 49
• The authors see good prospects for useful algorithmic development when efforts are directed by the needs of the situated decision-making task.They attribute the modest intervention’s practitioner engagement and praise to participatory design and deep analysis of situated work-as-done.Source: §5 · p. 49

Declarations

• B. Wilson acknowledges support from UK EPSRC grant EP/S021892/1, EMRTS Cymru, and the European Union.M. Roach acknowledges support from the European Union.Source: Funding · pp. 49–50

Ethics statement

• Ethical approval covered immersion and observations; EMRTS staff participants gave informed consent before taking part.The approvals included IRAS project IDs 312173 and 316886, and Confidentiality Advisory Group application 24CAG0106 for observations.Source: Ethics statement · p. 50

• BW is credited with conceptualisation, project administration, methodology, investigation, data curation, formal analysis, visualisation, and both forms of writing.Source: Contribution statement · p. 50
• MR and DR are credited with resources, supervision, funding acquisition, and writing-review and editing.Source: Contribution statement · p. 50
• GB is credited with resources and project administration, while CC is credited with resources.Source: Contribution statement · p. 50

• The stated purpose of applying the licence is Open Access.Source: Open Access · p. 50
• A CC BY licence was applied to Author Accepted Manuscript versions arising from the submission.Source: Open Access · p. 50
• The licence applies to any such Author Accepted Manuscript version arising from the submission.Source: Open Access · p. 50

• The authors report no generative AI use in preparing the manuscript text.Source: Use of Generative AI · p. 50
• Generative AI was used to create sketches from field photographs to avoid making individuals identifiable in figures.Source: Use of Generative AI · p. 50
• Spelling and grammar check tools were used during authoring.Source: Use of Generative AI · p. 50