LangGraph blueprint · Hospitals
Emergency triage & bed allocation
An agent suggests acuity, checks live capacity and drafts the handoff; nurses make every call.
- Steps
- 7
- Subjective questions
- 32
- Objective questions
- 32
The problem and the outcome
Problem: Crowded emergency departments lose precious minutes to manual triage and phone calls for beds.
Outcome: Faster triage and bed assignment, with clinicians deciding every step.
Architecture
How it works, step by step
- At arrival, vitals and symptoms are captured by a nurse or a kiosk.
- A triage agent suggests an acuity level (ESI 1–5) and shows its reasons.
- Critical patterns, such as stroke or sepsis signs, page the rapid-response team at once.
- A planner tool checks live bed, ventilator and nurse capacity across wards.
- It drafts a handoff summary for the receiving ward.
- The charge nurse confirms or overrides; nothing moves without a person.
- Every suggestion and override is logged for clinical audit.
Tech stack
- LangGraph
- Vitals and symptoms intake
- ESI triage guidelines retrieval
- Bed and staffing system APIs
- Paging integration
- Full audit trail
Key design decisions
- The agent suggests a triage level; nurses decide, and can always override.
- Hard rules for critical vital signs page the team immediately, without waiting for the model.
- Live capacity data is read from source systems, never estimated by the model.
Common mistakes
- Letting a model's suggestion delay care for a critical patient.
- Alert fatigue from too many low-value pages.
- No record of who accepted or changed each suggestion.
What to measure
- Door-to-triage time
- Agreement between suggested and final acuity
- Under-triage rate (most important)
- Time to bed
Most-asked interview questions
Subjective questions 32
1Describe the architecture of an AI-assisted triage and bed allocation system.Asked almost every time
Intake data (vitals, chief complaint, history) flows into deterministic rules for red-flag conditions and a model that suggests an acuity level with reasons. The suggestion is shown to the triage nurse, who decides. A separate optimiser recommends beds using real-time occupancy, staffing and isolation needs, with charge-nurse approval and full audit trails.
2How do you validate a triage model before clinical use?Asked almost every time
Retrospective evaluation on historical cases against expert triage and outcomes, subgroup analysis for bias, prospective shadow mode where the model runs silently alongside nurses, clinical governance review, and staged rollout with monitoring and a rollback plan.
3Why combine rules with a model?Asked almost every time
Rules guarantee that known critical patterns (chest pain with abnormal vitals, stroke signs, low oxygen) always escalate, independent of model errors. The model helps with nuanced cases; rules set a safety floor it can never lower.
4How do you design AI for a safety-critical setting such as an emergency department?Very often asked
Keep humans in charge of decisions, use deterministic rules for life-threatening signs, show the evidence behind each suggestion, log everything for audit, test heavily on historical cases with clinicians, and monitor under-triage closely. Roll out in shadow mode first, where suggestions are recorded but not shown.
5How do you check for bias in triage suggestions?Very often asked
Compare under- and over-triage rates across age, sex, ethnicity, language and insurance groups, investigate features that proxy for protected attributes, and adjust data or thresholds with clinical oversight.
6How should suggestions be presented to nurses?Very often asked
Show the suggested level with the top contributing factors and any red-flag rule triggers, clearly labelled as advisory, with one click to accept or override and a required reason for overrides above a threshold.
7How do you allocate beds fairly and safely?Very often asked
Model constraints (isolation, specialty unit, telemetry needs, gender policies, staffing ratios) and objectives (time to bed, acuity priority), solve with an optimiser, and present recommendations for human approval with reasons.
8How do you handle missing or wrong vital signs?Very often asked
Validate ranges at entry, flag implausible values for re-measurement, treat missing critical vitals as a reason to escalate rather than ignore, and never let the model fill them in.
9How do you monitor the system after go-live?Very often asked
Track agreement and override rates, under-triage events reviewed by clinicians, time to triage, bed wait times, alert fatigue, data drift and subgroup performance, with regular governance reviews.
10How do you keep the system available during outages?Very often asked
Design for graceful degradation: triage continues with standard manual protocols if the system is down, data entry is buffered, and the system never becomes a single point of failure for care.
11How do you avoid alert fatigue?Very often asked
Tune thresholds with clinicians, show only actionable alerts, group related alerts, measure acknowledgement and override rates, and retire alerts that add noise without benefit.
12How do you design the system for an emergency department surge, such as a mass-casualty event?Very often asked
Provide a surge mode that switches to a simplified mass-casualty triage protocol, prioritises speed over detail, raises bed and staff alerts to command staff, and suspends non-essential suggestions. Rehearse it in drills so staff know how the tool behaves.
13How do you combine nurse judgement and model suggestions in the record?Very often asked
Store the model's suggested level, its inputs and version, and the nurse's final level and reason separately. Reports can then show agreement rates and outcomes for both, without overwriting what the nurse decided.
14What is shadow mode and why use it?Often asked
The system runs on real cases, but its outputs are only recorded, not acted on. You compare them with what clinicians actually decided, measure accuracy and risk, and fix problems before anyone relies on the system.
15Why is under-triage worse than over-triage?Often asked
Under-triage means a seriously ill patient waits too long, which can cause harm or death. Over-triage uses extra resources but is safer. Systems are tuned to minimise under-triage, even at the cost of more over-triage.
16What regulations apply to clinical decision support?Often asked
Depending on country, software that influences clinical decisions may be a medical device (FDA, EU MDR, CDSCO in India), requiring risk classification, validation evidence, quality management and post-market surveillance.
17How do you explain a model's suggestion?Often asked
Show the main factors (for example low oxygen saturation and high heart rate) and similar historical patterns in plain clinical language, without implying certainty the model does not have.
18How do you integrate with hospital systems?Often asked
Read and write through the EHR's APIs (FHIR or HL7 v2 ADT messages for admissions, discharges and transfers), keep the bed board in sync, and respect existing workflows.
19How do you forecast bed demand?Often asked
Use historical arrivals, admissions and discharges by time of day, seasonality and local events to predict occupancy hours ahead, helping managers open beds or call staff early.
20What does shadow mode tell you that retrospective tests do not?Often asked
How the system behaves on live, messy data and real workflow timing, whether its suggestions would have helped, and which edge cases appear in practice, without affecting patients.
21How do you retrain safely?Often asked
Version datasets and models, validate on recent data and subgroups, review with clinical governance, deploy behind a flag with comparison to the previous model, and keep rollback ready.
22Who is accountable for triage decisions?Often asked
Clinicians and the hospital remain accountable; the system is advisory, and policies should say so, with overrides documented.
23How do you design the data pipeline for real-time use?Often asked
Stream vitals and registration events, compute features within seconds, keep a short latency budget for suggestions, and log every input and output for review.
24How do you handle paediatric patients differently?Often asked
Use age-specific vital sign ranges and paediatric triage criteria, validate the model separately on children, and route paediatric cases through rules that escalate on child-specific warning signs.
25How do you predict patient deterioration after triage?Often asked
Score early warning signals from repeated vitals and lab results over time, alert when the trend crosses a threshold, and re-prompt reassessment, since a patient's acuity can change while waiting.
26How do you plan transfers when beds are full?Often asked
Combine real-time bed status, expected discharges and forecast admissions, rank transfer options by clinical need and capability of receiving units or hospitals, and present them to bed managers with travel and handover requirements.
27How do you test the system's behaviour with incomplete integrations?Often asked
Simulate missing feeds such as no lab results or delayed vitals, confirm the tool shows clear warnings and falls back to rules instead of silently suggesting a lower acuity, and include these cases in the acceptance tests.
28What is under-triage and why is it dangerous?Sometimes asked
Assigning a lower urgency than needed, which delays care for patients who are seriously ill and can lead to harm or death; minimising it is the top priority.
29How do you involve clinicians in design?Sometimes asked
Co-design workflows with nurses and doctors, run usability tests in simulation, and keep a clinical champion group for feedback and governance.
30How do you handle language barriers at intake?Sometimes asked
Support interpreters and translated symptom forms, and avoid features that penalise patients who cannot describe symptoms in the local language.
31What should be logged for each suggestion?Sometimes asked
Inputs, model version, suggestion, contributing factors, rule triggers, nurse decision and override reason, with timestamps.
32How do you train staff to use the triage assistant well?Sometimes asked
Run short sessions with real anonymised cases, explain what the model sees and cannot see, show how and when to override, and gather feedback during the first weeks of live use.
Objective questions 32
1The final say on a patient's acuity belongs to…Asked almost every time
Why: The system is advisory only.
2Which error matters most to minimise?Asked almost every time
Why: Delayed care for the seriously ill.
3What guarantees critical patterns always escalate?Asked almost every time
Why: Rules set a safety floor.
4Who makes the final triage decision?Very often asked
Why: The agent assists; licensed staff decide.
5What is shadow mode?Very often asked
Why: Observe live behaviour safely.
6Missing critical vitals should…Very often asked
Why: Never guess clinical data.
7How is bias checked?Very often asked
Why: This is called subgroup analysis.
8What should overriding a suggestion require above a threshold?Very often asked
Why: Overrides feed review and learning.
9If the system goes down, triage should…Very often asked
Why: Never a single point of failure.
10Bed allocation constraints include…Very often asked
Why: Safety and policy limits.
11In a mass-casualty event the system should…Very often asked
Why: Speed and simplicity matter most in a surge.
12Model suggestions and nurse decisions should be stored…Very often asked
Why: Separate records allow honest comparison.
13How should critical vital signs be handled?Often asked
Why: Deterministic rules are faster and more reliable for life-threatening values.
14Which error matters most to minimise in triage?Often asked
Why: Under-triage delays care for patients who need it most.
15Clinical decision software may be regulated as…Often asked
Why: FDA, EU MDR or CDSCO.
16What reduces alert fatigue?Often asked
Why: Measure override rates.
17Which HL7 messages track admissions and transfers?Often asked
Why: Keep the bed board in sync.
18What helps managers open beds early?Often asked
Why: Predict demand hours ahead.
19How should model suggestions be explained?Often asked
Why: Avoid false certainty.
20How are new model versions deployed?Often asked
Why: Clinical governance first.
21Who remains accountable for triage?Often asked
Why: Policies should say so.
22What must be logged per suggestion?Often asked
Why: These logs are needed for review.
23Paediatric triage requires…Often asked
Why: Children's normal values differ from adults'.
24Deterioration alerts are based on…Often asked
Why: Acuity can change while patients wait.
25When a lab feed is missing, the tool should…Often asked
Why: Silent assumptions risk under-triage.
26Transfer options are ranked by…Often asked
Why: The right unit matters as much as an empty bed.
27Who should co-design the workflow?Sometimes asked
Why: Fit real clinical practice.
28Language barriers at intake are handled by…Sometimes asked
Why: Avoid penalising patients.
29Real-time suggestions need…Sometimes asked
Why: Triage happens in minutes.
30Implausible vital sign values should be…Sometimes asked
Why: Bad data leads to bad advice.
31Ongoing monitoring tracks…Sometimes asked
Why: Safety needs continuous review.
32Staff training should show…Sometimes asked
Why: Understanding limits builds safe trust.
Track your progress through every blueprint in your dashboard.
Start learning