Capital replacement committees operate under a recurring asymmetry. Evidence about acquisition cost is usually precise, whereas evidence about future downtime, failure consequences, vendor support, parts availability, residual life, and clinical substitution can be noisy or partly judgmental. Moreover, the costs of false decisions are asymmetric. Replacing a serviceable device too early wastes scarce capital and may trigger training, integration, and implementation costs. Deferring a genuinely unsafe or unsupported device can expose patients and staff to harm, extend downtime, or force emergency procurement under worse commercial conditions. This asymmetry motivates a decision architecture that is conservative about admissibility but explicit about uncertainty once safety constraints have been satisfied.

The paper focuses only on capital allocation for replacement of high-cost radiological equipment that uses ionizing radiation. It does not attempt to optimize preventive-maintenance frequency, corrective-repair dispatching, radiation-protection operations, procurement negotiation, cybersecurity operations, or the entire technology lifecycle. Those functions supply evidence or constraints to the replacement decision, but they are outside the optimization target. This narrower scope is deliberate: a reviewer should be able to identify the decision variable, the unit of analysis, and the proposed mechanism without treating the framework as a generic checklist for all biomedical engineering activity.

This article develops Risk-Gated Replacement Allocation (RGRA). The innovation is architectural rather than merely additive. RGRA separates three decisions that are often collapsed into one score: 1) safety admissibility, governed by non-compensatory gate conditions, 2) replacement priority, represented as an interval rather than a single deterministic rank, and 3) capital allocation, formulated as a budget-constrained portfolio choice. The separation is designed to make the decision process auditable: a device can be escalated because it fails a gate, prioritized because its robust interval dominates another device, or selected because its avoided expected lifecycle loss justifies scarce capital.

The paper contributes in four ways. First, it reframes safety as a feasibility condition rather than another weighted criterion. Second, it incorporates uncertainty in weights and evidence through robust priority intervals, reducing false confidence in precise ordinal rankings. Third, it distinguishes rank from allocation by introducing an explicit portfolio capital-allocation model. Fourth, it adds a value-of-information rule for borderline devices, directing measurement effort toward cases where additional evidence can actually change the decision. The main work is conceptual and design-oriented; no empirical performance claims are made without hospital data. Six falsifiable propositions and a prospective validation protocol are provided so that the framework can be tested rather than accepted as a managerial narrative.

The remainder of the article is organized as follows. Section 2 reviews replacement, reliability, maintenance prioritization, safety governance, and decision-analytic methods. Section 3 defines the conceptual design and variables. Section 4 presents the RGRA model and decision rules. Section 5 develops governance, implementation, and testable propositions. Section 6 concludes with limitations and future research.

Medical-equipment replacement research has long recognized that age alone is an inadequate criterion. Fennigkoh's model [3] used ten attributes spanning service and support, equipment function, cost benefits, and clinical efficacy. The subsequent Galliera Hospital comparison of a linear replacement model with fuzzy logic showed that model form materially affects discrimination among assets [4]. Huang et al. [5] combined multiple-criteria weighting and ranking methods for large-equipment replacement, reinforcing the importance of systematic rather than purely subjective decisions.

Park et al. [6] found occupational differences in replacement priorities and reported that safety was especially important to medical engineers. The implication is not that one set of weights is universally correct; it is that replacement scores encode organizational preferences and professional perspectives that may change across hospitals and device classes.

The broad health-technology-assessment literature reaches a similar conclusion. Pimenta and Vieira [7] reviewed 51 prioritization studies and found that multi-criteria methods were dominant among studies that specified their value-assessment method, but no generally accepted approach existed. This finding exposes a methodological risk for replacement committees: The apparent precision of a final score may be much greater than the precision of the underlying criteria, weights, and stakeholder judgments. A point rank of 1 versus 2 can therefore be unstable even when the reporting format looks definitive.

Lifecycle-data approaches partly address subjectivity by using observed maintenance histories. Liao et al. [8] used a discrete-time Markov chain and Markov decision process to study repair and replacement using 24,516 repair and maintenance records for 5,171 products.

Such models demonstrate the value of transition histories and lifecycle costs, but their objective is different from annual multi-device capital allocation. A hospital still needs a governance rule for combining technical evidence with clinical consequence and for handling devices that present safety conditions before an economic repair-replace comparison is meaningful.

Replacement decisions depend on evidence generated by maintenance systems, but maintenance prioritization should not be confused with replacement prioritization. Taghipour et al. [13] proposed a structured approach for prioritizing medical equipment for maintenance, establishing the importance of function, risk, and maintenance-related attributes.

Saleh et al. [14] developed a preventive-maintenance prioritization index using quality function deployment, while Jamshidi et al. [15] proposed a fuzzy risk-based framework for prioritizing medical devices. Hernandez-Lopez et al. [16] also developed an index for preventive-maintenance prioritization. These models are operationally valuable, yet their dependent decision is normally which maintenance activity deserves attention, not which capital replacement should be authorized.

Operations-research models further clarify the distinction. Ben Houria et al. [17] developed quantitative techniques for medical-equipment maintenance management, and Hamdi et al. [18] proposed intelligent work-order prioritization. These tools may generate important replacement signals such as repeated corrective work, downtime burden, maintenance complexity, and inability to restore performance.

However, a high maintenance priority need not imply immediate replacement: a newer critical device can require urgent repair yet still have strong supportability and remaining life. Conversely, an apparently functioning asset can merit replacement because of end-of-support, recurrent safety alerts, or inability to meet contemporary clinical requirements.

Reliability evidence has become increasingly data-driven. A systematic review by Zamzam et al. [19] synthesized approaches to medical-equipment reliability assessment. Abd Rahman et al. [20] used machine learning to predict device failure, and Zamzam et al. [21] developed integrated failure-analysis models for maintenance management. Ye et al. [22] reported a 2026 real-hospital deployment of machine-learning-assisted maintenance prioritization across 9,924 devices, demonstrating that predictive systems can be operationally embedded. These advances support better evidence for replacement decisions but do not eliminate the decision problem: a probability of failure must still be translated into consequence, intervention, and capital choices.

Recent replacement-oriented evidence reinforces the relevance of condition and supportability. Amran et al. [23] integrated reliability, replacement priority, and maintenance strategy criteria in Malaysian hospitals, including safety alerts, age, utilization, failure and availability indicators, downtime, alternatives, cost, and support-related factors.

Wang et al. [24] examined patient incidents associated with maintenance omissions, highlighting why missing maintenance actions can carry safety consequences. Evidence-based equipment-management implementations [25] and computerized maintenance systems in hospital settings [26] also show that data quality, inventory completeness, and standardized work-order recording are prerequisites for defensible lifecycle decisions. Accordingly, RGRA treats maintenance and reliability data as evidence inputs, not as substitutes for replacement governance.

Multi-criteria decision analysis provides formal tools for combining heterogeneous evidence, but different families of methods embody different assumptions. The Analytic Hierarchy Process developed by Saaty [27] structures pairwise judgments and derives relative weights.

Promethee, introduced by Brans and Vincke [28], and ELECTRE-family outranking methods associated with Roy [29] can support preference structures that do not reduce every comparison to a single additive utility. These methods are relevant because replacement choices often involve veto-like conditions: A severe safety or compliance concern should not be neutralized by low acquisition cost or good utilization performance. A second problem is uncertainty. Replacement scores depend on estimated failure rates, uncertain remaining life, subjective severity assessments, variable vendor support, and weights negotiated by different stakeholders. Robust optimization offers a disciplined alternative to pretending that uncertain inputs are known exactly. Bertsimas and Sim [30] formalized a tractable approach that controls the conservatism of solutions under data uncertainty, and Bertsimas et al. [31] reviewed the theory and applications of robust optimization. The key managerial idea adopted here is not a particular uncertainty set; it is the principle that a decision should remain defensible over a plausible set of inputs rather than only at one nominal estimate. The literature therefore suggests a three-part gap. Existing replacement models provide criteria and ranks; reliability models improve evidence; and decision science provides non-compensatory and robust tools. What is missing is a focused hospital replacement architecture that connects them in the correct order: Safety admissibility first, robust priority second, and capital allocation third. Table 1 summarizes this synthesis.

HTA prioritization [7] Broad evidence on value attributes, stakeholder involvement, and MCDA prevalence.

Lifecycle repair- Uses observed maintenance histories replace optimization and transition dynamics to optimize [8] repair or replacement.

Maintenance and Failure, downtime, maintenance reliability analytics [13- burden, condition, and predictive 26] evidence.

MCDA and Formal preference, trade-off, and veto outranking [27-29] structures.

Robust optimization Decision stability under uncertain [30], [31] parameters.

optimization requirements.

therefore begin with an admissibility screen.

increase failure, workflow, and safety exposure [39].

Remaining Problem for Replacement Governance

Typically returns a score or rank; severe safety conditions may remain compensable.

No generally accepted prioritization approach; context and weight choices remain material.

Does not by itself solve multi-device annual capital allocation with heterogeneous safety gates.

Predicts or prioritizes maintenance; does not define replacement admissibility and

This study is an integrative conceptual review and decision-model design, not an empirical effectiveness study. The evidence base was organized around five streams: medical-equipment replacement, maintenance and reliability, healthcare technology management, medical-device and radiation-safety governance, and decision analysis. The design criterion was explanatory parsimony: every construct included in RGRA must answer one of three decisions, admissibility, priority, or capital selection. Constructs that belong primarily to other managerial problems, such as preventive-maintenance interval optimization or corrective work-order dispatch, enter only as evidence variables.

The model deliberately distinguishes repair from capital renewal. Tarakci et al. [37] address which failed device should be repaired first, whereas RGRA asks which radiological systems should receive scarce replacement capital before or instead of further repair. WHO maintenance guidance [38] is used to define maintenance evidence and safe equipment management; ESR renewal guidance [39], CT obsolescence evidence [40], and radiotherapy capital analysis [41] define the narrower renewal context. Mixing these outcomes would recreate the broad managerial framework this article is designed to avoid.

The conceptual model assumes a hospital has an asset register and a minimum set of device-level evidence. A practical implementation can be built from a computerized maintenance management system, procurement and vendor records, safety-test documentation, incident or alert records, and clinical service information. Where evidence is missing, the model does not silently impute precision; uncertainty is widened and, when potentially decision-changing, a value-of-information review is triggered.

Let i = 1,...,N index candidate devices or homogeneous asset groups considered in one capital-planning cycle. The binary decision x_i equals 1 when replacement of candidate i is funded in the planning horizon and 0 otherwise. Candidate inclusion is itself governed: only assets meeting the hospital's eligibility rule - for example approaching end of supported life, repeated significant failures, persistent unavailability, material capability deficit, or documented safety concern - enter the replacement review. This prevents annual ranking of every low-risk asset merely because data exist.

For each candidate, the evidence vector z_i contains normalized indicators that are intentionally narrower than a general technology-management score: C_i for clinical consequence if unavailable or malfunctioning; R_i for reliability deterioration and recurrent failure; D_i for downtime and service-continuity exposure; O_i for technological obsolescence and supportability; Q_i for sub-threshold quality-assurance or dose-performance deterioration that has not itself triggered a safety gate; L_i for lifecycle burden of continued retention; and A_i for lack of redundant clinical capacity. These variables need not share identical uncertainty. Age can inform O_i or lifecycle modeling but is not an autonomous replacement objective: two equally old CT scanners can have very different detector performance, support status, failure histories, protocol capability, and service dependencies.

Safety-gate evidence is stored separately as g_im for gate condition m. This separation is essential. If g_im were simply another normalized criterion in z_i, the model would again permit economic or utilization benefits to compensate for a condition that should trigger escalation. The core variables and suggested operational sources are shown in Table 2.

set reproduces false precision.

rebuilding scores from memory.

model-validation procedures.

Illustrative Evidence Source Decision Role

Radiation QA; safety tests; Admissibility incident/alert records; regulator or manufacturer instructions.

Clinical service review; modality Priority dependency; case mix; critical pathway mapping.

CMMS failure history; service logs; Priority repeat-call and component records.

Downtime hours; work orders; Priority vendor response and parts lead times.

Vendor roadmaps; support notices; Priority parts availability; upgrade history.

Medical-physics QA; constancy Priority tests; protocol review; performance trending.

Finance; contracts; parts; energy; Value maintenance and commissioning plans.

Utilization; scheduling; installed Constraint capacity; referral alternatives.

Procurement quotation; facilities Capital plan; physics/engineering implementation plan.

RGRA does not require one universal uncertainty model. A hospital may use simple scenario bounds, expert-derived intervals, bootstrapped reliability estimates, or statistical prediction intervals. The minimum requirement is that uncertainty be visible. Let W denote a plausible set of criterion weights and Z_i a plausible set of evidence vectors for candidate i. Uncertainty sets can reflect inter-rater disagreement, measurement error, sparse maintenance history, or forecast uncertainty. Their size should be justified and version-controlled, because an arbitrarily wide set can make every decision ambiguous while an unrealistically narrow

Auditability requires recording four elements for each cycle: The evidence snapshot, the gate decision and rationale, the uncertainty assumptions used for robust ranking, and the optimization inputs that produced the funded set. This design allows later post-completion review to distinguish model error from data error, policy override, or new information. It also supports calibration across annual cycles rather than repeatedly

In accordance with the journal's authorship and artificial-intelligence disclosure policy, a large language model was used during manuscript development for structural drafting, language refinement, and organization of the conceptual framework. The author is responsible for verification of references, equations, claims, and final wording; the model is not an author and no empirical results were generated by artificial intelligence. Any future empirical implementation should separately document software, code, data provenance, and

The first layer determines whether ordinary retain-versus-replace trade-offs are admissible. Let tau_m denote the escalation threshold for gate condition m. A candidate fails the safety gate when at least one predefined condition crosses its threshold:

Gᵢ = { maxₘ(gᵢₘ − τₘ) ≥ 0}.

resolved.

admissibility layer.

Gate Condition Non-Compensation Rationale

Radiation QA or Dose, output, geometry, or protection condition protection requirements makes clinical use cannot be offset by lower unacceptable cost or high utilization.

Mandatory recurrent or An unresolved mandatory post-repair safety test is test failure changes the failed feasible action set.

Serious field-safety External safety information action remains can invalidate ordinary unresolved retain/defer choices.

Required performance Persistent inability to cannot be sustained maintain essential performance is not an economic trade-off.

Critical support loss End-of-support becomes a creates unacceptable gate only when safe residual risk operation cannot be assured with validated controls.

Immediate Re-Entry Condition Response

Restrict affected use; Condition corrected, medical-physics and verified, and formally radiation-safety released.

escalation.

Restrict use as required; Documented pass or investigate, correct, and approved disposition.

retest.

Apply Field action closed and manufacturer/regulatory residual risk accepted.

action; quarantine, correct, or replace.

Engineering, clinical, and Verified restoration of physics review; consider required performance.

withdrawal or replacement.

Escalate supportability; Validated support/control impose interim controls; restored or asset replaced.

accelerate replacement.

Here G_i = 1 denotes a gate-triggered case. The gate is not synonymous with automatic purchase. Its function is to remove routine deferral from the compensatory ranking process and trigger a governed response: Restrict or suspend use where required, correct the condition if feasible, escalate to the appropriate engineering/clinical/radiation-safety authority, or pursue replacement/decommissioning. A gate can therefore result in corrective action followed by re-entry into the ranking set if the underlying condition is

The most important design rule is that gate definitions must be local, explicit, and linked to authoritative requirements. Examples include a failed mandatory electrical-safety or post-repair test, an unresolved serious field-safety action, inability to maintain required safety performance, or a radiological quality-assurance condition that makes continued clinical use unacceptable. The gate should not include every undesirable attribute; high maintenance cost or low utilization, for example, belong in the priority model rather than the

Fig. 1 shows the complete RGRA architecture for radiological equipment. The principal distinction is the branch after the safety-admissibility stage: a gate-fail case is escalated before ordinary economic ranking, so a severe radiological or mandatory condition cannot disappear inside a weighted average.

Candidates that pass the gate are evaluated using a transparent replacement-priority function. A simple additive form is useful as a baseline because it is interpretable and easy to audit:

Pᵢ(w, zᵢ) = Σₖ₌₁ᴷ wₖzᵢₖ, wₖ ≥ 0, Σₖ wₖ = 1. (2)

The additive form is not claimed to be uniquely correct. Hospitals may use a validated MCDA or outranking approach. What RGRA requires is that the resulting priority be stress-tested under plausible uncertainty. For an additive implementation, define lower and upper robust priority bounds:

Pᵢ⁻ = min_{w∈W, zᵢ∈Zᵢ} Pᵢ(w,zᵢ), Pᵢ⁺ = max_{w∈W, zᵢ∈Zᵢ} Pᵢ(w,zᵢ). (3)

The interval [P_i^-, P_i^+] communicates two pieces of information that a point rank hides: The strength of a candidate's case and the stability of its ordering. If the interval for one device lies entirely above another device, the preference is robust to the modeled uncertainty. If intervals overlap materially, the rank should be treated as unstable rather than reported with false precision.

A planning threshold can also be expressed as an interval decision: devices whose lower bound exceeds a threshold have a strong replacement case; those whose upper bound remains below it have a weak case; and threshold-crossing candidates are natural targets for further evidence collection.

Fig. 2 is deliberately illustrative rather than empirical. It shows how robust intervals change managerial interpretation across radiological candidates.

CT-1 is clearly above the planning threshold; Radiography-1 is clearly below it; and Fluoroscopy-1 straddles the threshold. A point ranking would hide that distinction in confidence.

A ranking still does not solve the budget problem. Suppose a high-priority linear accelerator or CT system costs several times more than a high-priority radiography unit. Selecting the top N ranks ignores acquisition cost, installation cost, capacity constraints, and the total risk reduction that can be achieved by different combinations. RGRA therefore converts replacement priority into a value model. Let V_i denote the expected lifecycle loss avoided by replacing rather than retaining candidate i over the planning horizon:

Vᵢ = E(Lᵢ | retain) − E(Lᵢ | replace). (3)

Loss can include expected failure and downtime consequences, maintenance and service burden, emergency procurement exposure, and other locally approved components. Clinical harm should be monetized only where the hospital has a defensible framework; otherwise consequence terms can remain in a multi-objective or constrained formulation. The essential point is that V_i represents the decision-relevant advantage of replacement, not merely a rank score. The funded replacement set is then selected under a capital budget B:

maxₓ Σᵢ xᵢVᵢ subject to Σᵢ xᵢKᵢ ≤ B, xᵢ ∈ {0,1}. (5)

Additional constraints can represent service continuity, medical-physics commissioning capacity, installation dependencies, or minimum modality coverage. For example, a hospital can prohibit overlapping replacement outages when the remaining CT or treatment capacity would fall below a prespecified service floor. When avoided-loss values vary materially across plausible scenarios s in S, a conservative capital plan can maximize the worst-case portfolio value:

maxₓ min_{s∈S} Σᵢ xᵢVᵢˢ subject to Σᵢ xᵢKᵢˢ ≤ B, xᵢ ∈ {0,1}. (3)

Eq. (6) is not mandatory for every hospital; it is a robust counterpart for settings in which lifecycle value or implementation cost is sufficiently uncertain that a nominal knapsack can be unstable. RGRA therefore separates the ordinal question 'which device has the stronger replacement case?' from the allocation question 'which combination remains defensible within the actual capital plan across credible scenarios?'.

Fig. 3 provides a schematic capital-risk frontier at the portfolio level. Open circles on the dashed curve represent efficient feasible portfolios, grey circles represent dominated feasible alternatives, and the black circle marks the illustrative portfolio selected at the capital ceiling. The frontier emphasizes that a higher-capital portfolio is justified only when the associated reduction in residual lifecycle loss is sufficient; the vertical line is an illustrative budget boundary rather than empirical hospital data.

Uncertainty should sometimes trigger measurement rather than immediate replacement or deferral. Define a decision-relevant information set q for a borderline device, for example a detailed vendor support confirmation, calibrated performance test, clinical redundancy audit, or component-condition assessment. The Expected Value of Information (EVI) can be written conceptually as the expected improvement in the best attainable decision after information is obtained, net of the cost of acquiring that information. RGRA uses this principle selectively: Evidence collection is prioritized when it has a realistic probability of changing the gate status, interval classification, or funded portfolio.

EVI(q) = E_q[maxₐ U(a | q)] − maxₐ E[U(a)] − Cost(q). (7)

The value-of-information layer prevents two opposite errors. Without it, committees may repeatedly request more analysis even when the choice is already robust, delaying necessary replacement. Alternatively, they may commit capital to a borderline device because existing records are incomplete when one targeted inspection or vendor confirmation could resolve the uncertainty. Operationally, the rule is simple: collect more evidence when the expected decision benefit exceeds its cost and delay burden.

Implementation should begin with a narrow candidate register, not an enterprise-wide attempt to score every device. Each candidate record should contain an asset identifier, device class, clinical service, gate evidence, failure and downtime history, supportability status, replacement-cost estimate, redundancy information, and the version of the criteria and weights used. Data from a computerized maintenance management system should be joined to safety-test records, procurement/vendor data, and clinical-capacity information using controlled identifiers. Free-text service records can be reviewed, but high-stakes gate decisions should not depend solely on an opaque text classifier.

Decision rights should also be separated. Biomedical or clinical engineering validates technical and maintenance evidence; clinical service owners validate consequence and redundancy; procurement/finance validate replacement cost and implementation burden; medical physics or radiation-safety functions validate applicable radiological conditions; and the capital committee approves the final funded set. A model owner maintains criteria, uncertainty rules, and audit logs, but should not have unilateral authority to override a gate. Overrides, when legally and clinically permissible, should be explicitly documented with accountable approval. This governance structure directly addresses a common failure mode of decision-support systems: The score becomes the decision. In RGRA, the model is a structured argument. It shows what evidence triggered a gate, what uncertainty drives the priority interval, and why the optimization selected one combination over another. Human judgment remains necessary, but it becomes reviewable rather than hidden in an unexplained final ranking.

Because RGRA is conceptual, its contribution depends on whether its mechanisms can be falsified. The following propositions are therefore stated in comparative terms against realistic baselines such as age-only replacement, deterministic weighted scoring, or rank-first capital selection.

Proposition 1 (safety protection). A non-compensatory safety gate will reduce the rate at which devices with unresolved high-consequence safety conditions are retained or deferred relative to a purely additive replacement score, holding candidate information constant.

Proposition 2 (rank stability). Robust priority intervals will produce fewer decision reversals under plausible perturbations of criterion weights and measurement inputs than a single nominal weighted-score ranking. Proposition 3 (capital efficiency). A budget-constrained portfolio selection based on replacement value will achieve greater avoided expected lifecycle loss per unit of capital than selecting candidates solely by descending priority rank, when replacement costs are heterogeneous.

Proposition 4 (supportability interaction). The effect of obsolescence and supportability on replacement priority will be stronger when clinical redundancy is low and consequence of unavailability is high than when substitutes are readily available.

Proposition 5 (information targeting). Value-of-information-guided evidence collection will reduce unnecessary immediate replacement or unstructured deferral among threshold-crossing candidates compared with uniform additional review of all candidates.

Proposition 6 (closed-loop calibration). Updating uncertainty sets, thresholds, and evidence mappings using realized post-decision downtime, maintenance cost, service disruption, and safety outcomes will improve out-of-sample decision calibration relative to static age- or score-based rules.

protection safety as an ordinary rate for expert-adjudicated gate plus prospective gate review.

criterion. cases.

P2 rank stability Nominal fixed-weight Rank reversals; top-k overlap; Bootstrap/scenario perturbation ranking. threshold classification stability. of evidence and stakeholder weights.

P3 capital Replace top ranks Avoided lifecycle loss per Historical replay or prospective efficiency until budget is capital unit; service interruption parallel capital-planning exhausted. avoided. simulation.

P4 supportability Additive main- Marginal priority effect across Regression or hierarchical model interaction effects-only score. redundancy/consequence with prespecified interaction strata. terms.

P5 information Uniform extra review Decision changes per review Prospective randomized or targeting or immediate point- cost; avoidable delay; stepped-wedge review workflow score decision. unnecessary replacement. where feasible.

P6 Closed-loop Static Calibration of predicted loss Rolling-origin annual evaluation calibration criteria/thresholds. and decision outcomes across with locked rules per cycle.

cycles.

A credible validation should use historical replacement candidates and subsequent outcomes without leaking future information into the original decision date. At each historical planning cycle, the model should be reconstructed from records available at that time, then compared with actual committee decisions and alternative baselines. Candidate-level outcomes can include emergency replacement, high-severity failure, unplanned downtime, corrective-maintenance burden, and documented safety escalation. Portfolio-level outcomes can include capital used, planned-versus-emergency procurement, aggregate downtime, and concentration of risk across clinical services.

The strongest validation would be prospective. RGRA recommendations could initially run in shadow mode while the hospital retains its existing process. Disagreements between RGRA and the committee would be adjudicated without forcing action, generating a structured error taxonomy. After calibration, a phased or stepped implementation could compare services or planning cycles while preserving safety oversight. Statistical evaluation should report uncertainty around outcome differences rather than relying on a single success rate. Sensitivity to weight sets, missing data, and alternative loss definitions should be prespecified. The model should also be evaluated for burden and interpretability. A replacement framework that requires months of manual data collection for every device may be academically elegant but operationally unusable. Metrics should therefore include analyst time per candidate, proportion of evidence automatically available, number of committee overrides, reason for override, and time from candidate nomination to decision. These implementation outcomes matter because the practical value of decision support is partly determined by whether it improves decision quality without creating an unsustainable administrative process.

RGRA is not appropriate for every equipment decision. When a regulator, manufacturer, or governing safety authority mandates withdrawal, replacement, or corrective action, optimization is secondary to compliance. When only one candidate exists and its replacement is mandatory, portfolio optimization adds no value. Conversely, for low-risk commodity equipment with abundant substitutes, a simplified lifecycle-cost rule may be sufficient. The framework is most useful when multiple clinically important candidates compete for scarce capital and evidence contains both hard constraints and uncertain trade-offs.

Several failure modes require monitoring. First, poorly designed gates can become so broad that every aging device escalates, defeating prioritization. Second, narrow uncertainty sets can make robust intervals cosmetically precise; excessively broad sets can make them uninformative. Third, estimates of avoided lifecycle loss can embed double counting if failure, downtime, and maintenance cost are not carefully separated. Fourth, stakeholder weights can drift toward political bargaining if elicitation is not governed. Finally, missing CMMS data can systematically disadvantage departments with weaker documentation. Governance should therefore treat data completeness and model calibration as quality indicators rather than hidden technical details.