Chapter 10 — Three worked engagements¶
Three engagements arrive in the same month at a European healthcare-operations practice. One is framed as an ML mortality score to help prioritise ICU admission during surge. One is framed as build us an AI nurse scheduler. One is framed as AI for radiology — we need faster turnaround on critical findings. Each is plausible. Each has a sponsor with a budget, a vendor with a deck, and a timeline pulled from somewhere upstream of the clinical team.
This closing chapter runs each of the three through the method. The first drops at Tier 0 on the Abolition check and never reaches G2. The second passes Tier 0, clears G1 and G2, and routes at G3 to Non-AI automation — a classical operations-research solver, not an AI system. The third decomposes into two pieces at G2 and routes to a mix of AI as assistant and Autonomous AI at G3, with the autonomy earned by running the assistant controls to maturity first. The trio is not a taxonomy and not a spectrum. It is what the method's outputs look like when the gates do their work.
10.1 Case A — ICU triage scoring during surge¶
The brief arrives from the medical directorate of a composite mid-sized German teaching hospital, roughly 900 beds, with an ICU of about forty. Two winters of respiratory-virus surge have left the unit repeatedly over capacity, and the directorate has read a vendor brochure proposing an ML score that ingests prior morbidity, age, a SOFA-style physiological summary, and a clinical-frailty scale, and returns a ranked triage priority. The stated goal is to help the ICU consultant on take prioritise admissions when the unit is full. The brochure emphasises that the score is advisory and that the consultant retains the decision.
Tier 0, first check — Technochauvinism. The check runs first because it is the cheaper of the two. The directorate's framing treats surge capacity as a prioritisation problem, which would be a technical problem if the prioritisation surface were the bottleneck. It is not. The bottleneck, in the G1-style framing conversation held with the ICU clinical lead, the nursing director, and the bed-management team, is bed availability, nursing ratios, and the discharge pipeline from ICU into step-down wards — the same set of operational constraints the unit has faced for a decade. An ML score does not add a bed, does not hire a nurse, and does not speed a step-down discharge. But the check still passes: there is a genuine clinical-decision artefact (which of two patients arriving within the same hour should be admitted first, given one slot), and a model could in principle contribute to it. The remainder, if AI were removed from scope, is a recognisable operational problem — but the brief is not solely an operational problem renamed. The check does not fire.
Tier 0, second check — Abolition. This one fires. The Abolition check is categorical and runs along four axes [1].
Dignitary harm. The score's output, rendered as a ranking, communicates a comparative claim about two patients' standing — that one should receive the scarce resource and the other should not, at a moment when the allocation decides whether the other patient survives. The mechanism of communication matters. A human triage protocol communicates the decision as a judgement under explicit guidelines that the affected party or their family can later read, contest, and understand as such. A model score communicates it as a number with a confidence interval. The dignitary axis fails not because the number is wrong but because the form is wrong for a decision of this kind.
Irreversibility. When the score is wrong, the patient denied the slot dies or is permanently harmed. There is no review window that recovers the outcome. The consequence-time of a wrong call is hours; the rollback substrate would be the consultant's judgement, which was the substrate the score was meant to supplement. An irreversible decision with a short consequence time and no workable rollback fails the axis.
Due process. The German Federal Constitutional Court ruled in December 2021, in the case known as 1 BvR 1541/20, that the legislator must take effective measures ensuring that people with disabilities are not disadvantaged in the allocation of scarce life-sustaining resources in triage situations [2]. The court's reasoning attaches specifically to anti-discrimination protection: any triage criterion that systematically disadvantages a protected group produces a state duty to regulate. The empirical literature on SOFA and related severity-scoring systems, which the proposed model uses as features, documents exactly this disadvantage along racial and ethnic lines — the scores overpredict mortality for Black and Hispanic patients [3], which, inside a crisis-standards-of-care framework, systematically diverts resources away from those patients. A large fraction of Black patients in a JAMA Network Open simulation would have been improperly excluded from the highest-priority category by SOFA-driven allocation [4]. A system whose features are documented to discriminate against a protected group, deployed for a decision the court has named as triggering a duty of anti-discrimination protection, is a system whose reasoning cannot be meaningfully contested by the affected party, because the discriminatory pattern is inside the feature definitions.
Coercion. The patient did not consent to being scored and cannot opt out. The asymmetry between the operator's freedom to deploy and the patient's inability to decline is the point, and no interface design softens it.
Three of the four axes fire decisively. One is sufficient. The determination is categorical.
The jurisdictional layer reinforces, but does not drive, the determination. Annex III of the EU AI Act (Regulation 2024/1689) lists emergency-healthcare triage systems as high-risk, carrying the full set of compliance obligations for providers and deployers [5]. High-risk is not the same as off-limits; the Abolition refusal is not a jurisdictional refusal. The AI Act permits such systems to be marketed if they meet its controls regime. The method's refusal here is stronger: the controls regime does not reach the dignitary and due-process axes, because those axes are about the form of the decision, not the quality of the system.
See Illustration 10.1 for the determination's shape.
Illustration 10.1 — Case A's Tier 0 shape. The Technochauvinism check passes; the Abolition check fires decisively on dignitary, irreversibility, and due-process axes. The determination does not depend on future model capability or on a better-calibrated score.
The artefact. The Tier 0 determination, written as Chapter 4 specifies, reads:
Off-limits to AI, because dignitary, irreversibility, and due-process axes all fail. The decision to allocate a scarce life-sustaining resource communicates a claim about the patient's standing that a model score cannot defensibly carry; the wrong call is irreversible with no workable rollback substrate; the feature set is documented in the clinical literature to produce discriminatory error along protected categories, triggering the anti-discrimination duty recognised by the German Federal Constitutional Court in 1 BvR 1541/20. This finding does not depend on future model capability.
What the hospital should do instead. The refusal is not a refusal to engage. The remaining work, removed of AI scope, is a recognisable operational and policy programme: investing in step-down bed capacity to unblock ICU discharge; updating the written human triage protocol to be explicit about anti-discrimination protection, as the court requires; training the consultants on take in that protocol; and strengthening the surge-capacity planning that keeps the unit out of crisis triage in the first place. None of this is technical, and none of it needs AI. The vendor's proposal was, at root, a request to substitute a model for the policy and capacity work that the unit had not done. The Abolition refusal names that substitution and declines it.
10.2 Case B — Nurse rostering¶
The brief arrives from the chief nurse of a Nordic public teaching hospital, roughly 700 beds across five sites, as a sentence in an email: we need an AI scheduler for the ward-nursing roster, the current system is costing us agency spend and nobody likes the rotas. The sponsor's framing has AI in it. The method's first question is whether that framing is load-bearing.
Tier 0. Both checks pass. Technochauvinism is the more interesting of the two here. The current agency-spend problem has non-technical roots — a vacancy rate of about 9% across wards, a discharge pipeline that lengthens shifts unpredictably, and union-negotiated rest patterns that some managers had been informally overriding on paper rosters. Those are not problems a scheduler solves. But a separable technical problem does remain: given a fixed set of staff, certifications, union rules, and ward demand profiles, produce a monthly roster that satisfies constraints and minimises shift-swap churn. That is a genuine computational problem, and the check's remove-AI-from-scope rehearsal leaves it intact. Abolition is quiet: rostering decisions about working hours are subject to labour law and consultation, but they are not categorically off-limits to automation, and an automated roster is routinely the legitimate output of a scheduling system.
G1 Observe. The situation statement, drafted over two weeks of operator shadowing and conversations with the three senior sisters who currently draft the rosters, reads: the rostering function produces an acceptable roster each month, but the production cost is three full working weeks of senior-nurse time distributed across five sites; the current spreadsheet-plus-negotiation process cannot explore alternatives, which produces rosters that satisfy hard constraints but leave soft preferences on the table; last-minute vacancies are filled from an agency pool at a premium; union-rule breaches surface on audit two or three times a year. The stakeholder map names five: the senior sisters (whose time is consumed), the ward nurses (who receive rotas), the union representatives (who certify compliance), the finance director (who watches agency spend), and the workforce-systems team (who own the eventual software). No stakeholder in the map is calling the problem AI; the sponsor's framing came from a vendor meeting.
G2 Decompose. The issue tree opens on the rostering process and splits into three pieces: (i) the monthly roster-draft production itself, (ii) the mid-cycle adjustment for short-notice absences, and (iii) the audit-and-correction cycle that fires when a union-rule breach is discovered. Pieces (ii) and (iii) are downstream of (i) and arguably easier — (ii) is an operational routine, (iii) is an incident workflow. The sponsor's AI scheduler request is about piece (i). The other two pieces get their own rows on the eventual routing map. The rest of this section traces piece (i).
G3 Route, Q1 — does any automation earn its place? The value question passes cleanly. Three full working weeks of senior-nurse time each month, repeated twelve times a year across five sites, is high-volume. The decision has stable inputs (staff list, certifications, ward demand forecasts, union rules) and a measurable objective (constraint satisfaction plus a weighted sum of preference violations). A value-stream mapping exercise on the current flow identifies the roster-drafting step as the single biggest consumer of senior-nurse hours that is not direct care or direct supervision. Automation earns its place.
G3 Route, Q2 — does a rule or script fit? The decision surface is the interesting one. Rostering problems of this shape have been the canonical case study in operations research for thirty years [6]. The constraints are enumerable: hard constraints (certifications, legal rest, minimum ratios, contracted hours) and soft constraints (weekend rotation fairness, preference for consecutive early shifts, avoidance of isolated shifts, team continuity). The decision tables are large — hundreds of constraints across a hundred-and-fifty-nurse ward unit — but they are enumerable and they are stable, shifting only when union rules or staffing ratios change, which happens on the scale of years, not weeks. The surface is exactly the one Q2 is asking about.
An academic survey of the nurse-rostering literature documents the state of the art: mixed-integer programming, constraint programming, branch-and-price, and hybrid approaches including simulated annealing, with standardised benchmark instances maintained for comparison [6, 7]. Commercial products deployed across European hospitals — RLDatix Allocate [8], Qgenda, and the Netherlands-based ORTEC Workforce Scheduling [9] — carry the OR solver inside them. The word AI has recently been added to some of these products' marketing copy, usually in reference to demand-forecasting overlays or self-scheduling suggestion engines that sit alongside the solver. The solver itself is not AI. It is a deterministic optimiser over a declaratively specified constraint set.
The check's outcome is clear: piece (i) is routed to Non-AI automation. The substrate is a constraint-programming or mixed-integer-programming solver — an inspectable, deterministic optimiser — not a machine-learning model. The Non-AI / AI divider in this book sits between Rule and Classical ML; the solver is on the Non-AI side by construction, because it does not learn from data. It derives its rostering output from a specification.
See Illustration 10.2 for the per-piece routing map entry.
Illustration 10.2 — Case B's routing. Q1 and Q2 both pass; the piece resolves to a deterministic solver. The solver's audit trail is the constraint specification itself, which is why the route lands on the Non-AI side of the divider.
Controls sketch for Non-AI automation. Non-AI automation is not a zero-controls outcome. Chapter 7 makes the point that every G3 outcome carries a controls burden, and the solver's burden is specific. The specification is under change control: any amendment to a constraint is reviewed by the senior nurses and, for hard constraints, by the union representative. The solver's monthly output is checked for infeasibility and for soft-constraint violations, with a weekly cadence for the first three months. A quarterly fitness review confirms that the constraint set still reflects current union rules and ward demand profiles; the review is the trigger for re-specification, not a routine rubber stamp. A sunset criterion is declared: if the solver's rosters require manual override on more than 15% of nurse-shifts for any two consecutive months, the engagement retires and the specification is re-scoped.
Why this is not a consolation prize. It is important that Non-AI automation is a first-class outcome of G3, not a fallback. The sponsor's original framing — build us an AI scheduler — was an accident of vendor language, and running the gates corrects it. A classical OR solver at this piece of work is more auditable than any ML system, easier to extend when a constraint changes, and does not carry the label-budget obligation an ML route would have had. A ward-staffing audit can trace any given assignment back to the specific constraints that produced it, which is exactly the property the nursing director needs when the union representative asks how a particular rota was drawn. The method's value here is the routing discipline: the sponsor asked for AI; the gates found the work and placed it on the Non-AI side of the divider, and that placement is a stronger engagement, not a weaker one.
Piece (ii) and piece (iii), briefly. Piece (ii), short-notice absence adjustment, routes to AI as assistant at G3: an LLM-surfaced draft of swap proposals based on the constraint specification and the current roster, which the duty charge-nurse reviews and sends. The decision is judgement-heavy (who on the call-back list is actually willing and rested), but the option-generation is language-shaped and repetitive. Piece (iii), the audit-and-correction cycle, stays Human-operated; it is a low-frequency, high-consequence incident workflow that does not earn automation on volume. The routing map for the engagement, in Chapter 7's format, carries three rows: piece (i) as Non-AI automation, piece (ii) as AI as assistant, piece (iii) as Human-operated. Sequencing and commitment follow the patterns of Chapter 8 and Chapter 9.
10.3 Case C — Radiology critical-finding triage¶
The brief arrives from the head of radiology at a composite large NHS teaching trust: we need AI for radiology, specifically for acute stroke and intracranial haemorrhage, our turnaround on time-critical findings is too slow. The sponsor has attended a vendor demo at a major European radiology congress and can name three products. The trust has a framework route to procure from the national AI Diagnostic Fund [10], which makes deployment faster than a first-principles build. The case is typical of what AI-for-radiology engagements look like when they are well-shaped.
Tier 0. Both checks pass. Technochauvinism passes because the bottleneck is genuinely pattern-recognition at volume in a stable image modality where published evidence for AI triage support is substantial [11, 12, 13]. Abolition is the more careful check. Radiological findings are clinical decisions with real consequences; the question is whether the form of the automation fails one of the four axes. It does not, provided the automation is shaped as assist or as notification under controls. An autonomous diagnostic call that bypasses the radiologist entirely would approach the Abolition axes; the routes this engagement considers do not.
G1 Observe. The situation statement reads: the neuroradiology worklist receives about 140 non-contrast head CT scans per 24 hours, of which roughly 6% contain a time-critical finding — acute ischaemic stroke signs, intracranial haemorrhage, or related pathology — and for which door-to-treatment time is a load-bearing outcome; current turnaround from acquisition to reporting radiologist's first read averages 42 minutes for in-hours scans and 68 minutes out-of-hours, with a long tail in both cases; the clinical consequence of the tail is the population the pathway is most meant to serve. The stakeholder map: the reporting radiologists, the stroke team, the ED consultants who own the referral, the radiology IT team who own PACS integration, the medical director, and the trust information-governance lead. A clinical-safety officer is nominated at this stage for the subsequent gates.
G2 Decompose. The issue tree opens on the turnaround-time problem and resolves into a small set of pieces. The two that matter for this chapter are:
- Piece (i) — worklist reprioritisation. A model reads each incoming scan and, if it flags a likely-urgent finding, suggests that the scan be moved to the top of the reporting radiologist's worklist. The radiologist confirms the reprioritisation or dismisses it when they next pick up the worklist.
- Piece (ii) — auto-escalation for very-high-confidence critical findings. When a scan is flagged with high model confidence as a critical finding above a calibrated threshold — most clearly, a large intracranial haemorrhage — an immediate page is sent to the on-call stroke or neurosurgical team, with a deep link to the scan, simultaneously with the radiologist being notified. The escalation does not wait for the radiologist's first read.
Two further pieces are named on the tree and carry their own routing elsewhere (routine worklist management and the subsequent structured report drafting). The focus below is the (i)–(ii) pair, because that pair is where the method's AI as assistant and Autonomous AI outcomes show their working.
G3 Route.
Piece (i) — AI as assistant. The AI level is classical ML at the model layer, wrapped as a triage-flagging workflow. Q3's two-part question — does AI work, and can fitting controls be drawn — answers yes and yes. Published evidence from deployed products is substantial. Qure.ai's qER detects intracranial haemorrhage on non-contrast head CT at 97% sensitivity in post-market surveillance and is deployed across NHS hospital trusts under the AI in Health and Care Awards, with an early deployment at NHS Greater Glasgow and Clyde [11]. Brainomix 360 Stroke, including the e-ASPECTS component, is CE-marked and has been evaluated across 26 NHS hospitals and more than 80,000 patients, with published associations of improved thrombectomy rates and reduced door-in-door-out times [12, 14]. Aidoc's ICH algorithm is deployed across hundreds of European hospitals and is FDA-cleared and CE-marked for intracranial-haemorrhage triage [13]. The empirical fitness of the model class on this decision distribution is not in question. What is in question, for piece (i), is whether AI holds the decision or assists a human who holds it. The decision here is: read this scan now, not later. It is a scheduling decision with a direct clinical consequence. But the clinical decision — is there in fact a haemorrhage, and what is the next step — remains with the radiologist, who reads the scan in a reordered queue. That placement is AI as assistant. The radiologist's override is one click away, and the draft of the reprioritisation suggestion is presented as exactly that — a draft, not an action.
Piece (ii) — Autonomous AI, under controls. Piece (ii) is autonomous because the page fires without a human in the loop between the model and the on-call clinician. The AI's output is an action (page + link), not a suggestion. The question Chapter 7 asks here is the harder one: can the controls this decision needs be drawn? The answer has to include: a calibrated confidence threshold for firing (tuned so that the page's false-positive rate stays below a declared level, published in the commitment page), a rollback substrate (the non-autonomous path continues in parallel; if the page channel fails, the radiologist-read path still reports), a named owner with authority to suspend the autonomous channel within hours of a degradation, rollback triggers that fire on numeric conditions, a review cadence, and sunset criteria. The decision has short consequence time (the page reaches the on-call team within seconds), moderate blast radius (one patient per page, but a bad page cascades into a cancelled or delayed response for the next real case), and is partly reversible (a false page interrupts the clinician but does not harm the patient; a missed page is bounded because the radiologist still reads the scan in the ordinary flow). The controls design fits the decision's shape.
G4 Sequence. The dependency graph is narrow but consequential. Piece (i) runs first — it is the less autonomous route, the more reversible, and the lower blast radius. Piece (ii) does not run in parallel. Piece (ii) runs only after piece (i) has operated in production for a declared period (the engagement sets this at three months) and cleared a controls-maturity gate: the assistant-mode override rate is stable below a threshold, the model's on-site calibration on the trust's own data has been verified, the radiologist team has signed off on the flagging behaviour's shape, the rollback-to-manual path has been tested in staging, and the on-call paging substrate has been tested for piece (ii)'s eventual use. This is the book's assist-before-autonomous pattern at its cleanest. Piece (ii) does not earn its autonomy from the model's research-paper performance; it earns it from having watched piece (i) in production long enough to know what the same model fails on at this specific trust.
G5 Commit — piece (ii)'s commitment page sketch. The page is one side of A4 and has the fields Chapter 9 specifies. For the worked sketch:
Piece: Auto-escalation for very-high-confidence critical-finding flagging on non-contrast head CT. Autonomous AI.
Owner: The clinical-safety officer for neuroradiology, named by person, signatory on the page. Authority to suspend the autonomous channel without approval. Default visibility into the dashboards listed below.
Rollback triggers. (a) False-positive page rate above 3% for any 14-day window → suspend autonomous channel; revert to the assistant-mode reprioritisation only, and the radiologist's ordinary reporting workflow carries the criticality signal. (b) Missed critical finding (a flagged-as-negative case later confirmed critical on radiologist read) rate above the pre-agreed baseline for any 14-day window → suspend the autonomous channel and convene a review. (c) Page-delivery latency above 90 seconds on any single shift → investigate; repeated for a week → suspend.
Review cadence. Weekly during the first quarter; fortnightly during the second; monthly thereafter. Chaired by the owner, attended by the radiology lead, the stroke-team lead, and the IT-integrations lead. A standing agenda: calibration report, override log, page-delivery log, incident review.
Sunset criteria. (a) Operator adoption below a defined threshold (measured as the rate at which pages lead to action within the target window) for any three-month period → retire. (b) Model calibration drift on trust-local data above a defined threshold on any monitored slice → retire and re-scope. (c) Published non-AI pathway performance (the assistant-mode alone, or the pre-existing manual path) reaches parity with the autonomous channel on time-to-treatment for any six-month window → retire; the additional autonomy is no longer adding value.
Signature line. The owner signs and dates. The engagement does not cross G5 without that signature.
See Illustration 10.3 for how piece (i) and piece (ii) sit relative to each other in the engagement's shape.
Illustration 10.3 — Case C's shape. Piece (i) runs in assistant mode first. Piece (ii) does not go live until piece (i) has cleared a controls-maturity gate: the autonomy is earned from operational evidence, not granted from research-paper performance. The dashed gate is the mechanism that makes the autonomy defensible.
What the engagement produces. The radiology department carries two live systems, a named clinical-safety-officer owner, a monthly review that both pieces are on the agenda of, and a set of numeric triggers and sunsets that are visible on a dashboard the owner sees by default. The vendors' products underneath are real and proven [11, 12, 13]; the method's contribution is not the model but the routing and the controls discipline that make Autonomous AI a defensible outcome rather than a leap.
10.4 What the trio shows¶
Four observations close the book.
Refusals are legitimate outputs. Case A produces a one-paragraph determination and stops. The method does not consider that a failure — it considers it the engagement's correct outcome, and the hospital's remaining work is recognisable as the operational and policy programme it always was. A practice that cannot produce a refusal when the situation calls for one is not a practice; it is a salesforce. The Abolition check is the refusal mechanism for categorical off-limits decisions; the Technochauvinism check is the refusal mechanism for problems that are not technical; the G3 outcomes Human-operated and Non-AI automation are quiet refusals of their own kind. Case A uses the first. Case B's piece (i) uses the fourth. Both refuse AI against the specific decision at hand. Both are positive outcomes of the method.
Most problems don't route to AI. The six outcomes of the method — two Tier-0 refusals plus four G3 placements — include four that are not AI. Case A lands on the Tier-0 refusal. Case B's piece (i) lands on Non-AI automation. The book's claim is not that AI is rarely useful; it is that most of the work in a serious engagement is routing, and routing often lands somewhere other than AI. When it does, the engagement is stronger for having arrived there through the gates rather than having defaulted to the sponsor's original framing. The sponsor of Case B asked for an AI scheduler and received a better-shaped engagement that places an OR solver at the right piece, an AI assistant at a different piece, and a human at a third. That is more, not less.
Assist-before-autonomous is what keeps autonomy safe. Case C's piece (ii) is autonomous, and the autonomy is defensible. It is defensible because piece (i) ran first for three months in assistant mode; because a named owner has authority to suspend piece (ii) within hours of a degradation; because the rollback substrate is the same assistant-mode path that carried the decision during the preparatory quarter; because the triggers fire on numeric conditions and the sunsets are declared in writing before launch. Remove any one of those and the autonomy is no longer defensible. The sequence is not a best practice — it is the mechanism by which autonomy becomes reviewable. Autonomy that skips the assistant-mode phase is not faster; it is more fragile, and the fragility catches up at the first incident.
The method's value is defensible routing, not using AI. The three cases each close the book on the same point: the engagement's output is a determination, a routing map, a sequence, and a commitment, each of which is specific enough to review. The method does not promise better AI. It promises a stronger engagement by a margin that a sceptical reader can trace through the gates. The toolkit — forty-five cards covering Tier 0 and all five Tier-1 gates — is where the individual tools live; a chapter of prose has pointed to them throughout. When a gate turns up on an engagement you are running, the card for that gate's tool is the smaller artefact to pick up. The book's last practical advice is that one: the gates are what the engagement runs; the cards are what the gates run on.
Sources¶
[1] McQuillan D. Resisting AI: An Anti-fascist Approach to Artificial Intelligence. Bristol University Press; 2022. [verified]
[2] Bundesverfassungsgericht. Order of the First Senate of 16 December 2021, 1 BvR 1541/20. The legislator must take effective measures ensuring that persons with disabilities are protected in triage situations caused by the pandemic. Available at: bundesverfassungsgericht.de/SharedDocs/Entscheidungen/EN/2021/12/rs20211216_1bvr154120en.html. [verified]
[3] Miller WD, Han X, Peek ME, Charan Ashana D, Parker WF. Accuracy of the Sequential Organ Failure Assessment Score for In-Hospital Mortality by Race and Relevance to Crisis Standards of Care. JAMA Network Open. 2021;4(6):e2113891. DOI:10.1001/jamanetworkopen.2021.13891. [verified]
[4] Ashana DC, Anesi GL, Liu VX, et al. Equitably Allocating Resources During Crises: Racial Differences in Mortality Prediction Models. American Journal of Respiratory and Critical Care Medicine. 2021;204(2):178–186. [verified]
[5] European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Annex III. Official Journal of the European Union; 2024. Annex III lists emergency-healthcare triage systems among high-risk AI systems. [verified]
[6] Burke EK, De Causmaecker P, Vanden Berghe G, Van Landeghem H. The state of the art of nurse rostering. Journal of Scheduling. 2004;7(6):441–499. [verified]
[7] Ceschia S, Dang NTT, De Causmaecker P, Haspeslagh S, Schaerf A. The Second International Nurse Rostering Competition. Annals of Operations Research. 2019;274(1):171–186. [verified]
[8] RLDatix (Allocate Software). Rostering product pages and NHS Workforce Alliance supplier profile. Available at: rldatix.com and workforcealliance.nhs.uk/suppliers/rldatix-allocate-software. [secondary]
[9] ORTEC. Workforce Scheduling in Healthcare — product documentation and Netherlands hospital case studies (Albert Schweitzer Hospital; Medisch Spectrum Twente). Available at: ortec.com/en/industries/healthcare. [secondary]
[10] UK Government, Department of Health and Social Care. £21 million to roll out artificial intelligence across the NHS (AI Diagnostic Fund announcement, June 2023). Available at: gov.uk/government/news/21-million-to-roll-out-artificial-intelligence-across-the-nhs. [verified]
[11] Qure.ai. qER — AI triage and notification for non-contrast head CT; deployment at NHS Greater Glasgow and Clyde under the NHS AI in Health and Care Award (ACcEPT project). Available at: qure.ai/product/qer and qure.ai/news_press_coverages. [secondary]
[12] Brainomix. Brainomix 360 Stroke, including e-ASPECTS — CE-marked decision-support software for acute stroke, with NHS evaluations across 26 hospitals. Available at: brainomix.com/stroke and the Health Innovation Network case study. [secondary]
[13] Aidoc. ICH (BriefCase) and stroke algorithms — FDA-cleared and CE-marked, deployed across European hospital networks (including Antwerp University Hospital and Isala Hospital, Netherlands). Available at: aidoc.com and radiology.healthairegister.com/products/aidoc-ich. [secondary]
[14] UK Government, Department of Health and Social Care. Artificial intelligence revolutionising NHS stroke care (AI stroke tools rolled out to 100% of stroke networks). Available at: gov.uk/government/news/artificial-intelligence-revolutionising-nhs-stroke-care. [verified]