4.1 The numerator problem

Volume and wRVUs quantify countable output, but the numerator excludes much of the work that makes imaging clinically useful: protocol selection, reviewing priors, consulting with referring teams, managing contrast risk, supervising trainees, communicating urgent findings, multidisciplinary conferences, protocol improvement, and resolving discrepancies.

If these are omitted, leaders can inadvertently reward their displacement. The result is apparent productivity accompanied by slower clinical closure, more interruptions, less teaching, or greater downstream rework.

4.2 The denominator problem

Per-day and per-shift measures assume scheduled time is the relevant denominator. The scarcest denominator may instead be subspecialty attention, technologist skill, nursing support, time in a functioning magnet, inpatient transport, or completed authorization.

Improving output per radiologist-day cannot release a scanner constraint, and extending scanner hours cannot resolve unavailable sedation support. A different model must identify which denominator is currently limiting the clinical aim.

4.3 The distribution problem

Averages conceal the tails that often determine resilience. If a small number of clinicians carry a large share of examinations, the service may appear productive while becoming highly sensitive to illness, resignation, retirement, or a schedule change.

Concentration should be measured directly through percentile ratios, Gini-like measures, and the proportion of output attributable to the top decile or quartile. These are descriptive, not moral judgments.

Why the mean cannot show you this

Between 2017 and 2024 the average examinations per radiologist-day moved from 49.1 to 49.4. A monthly report built on that mean would have shown a flat, healthy service across seven years. Over the same period the top quartile moved from 56.6 to 73.9 and worked 19.7% more clinical days.

Aggregate productivity can remain acceptable while the service becomes dependent on a small group working at levels that are difficult to reproduce or sustain.

Concentration is not automatically a problem

Expertise and specialization produce value. Concentration becomes a leadership risk when the dependency is invisible, unsupported, or irreplaceable.

Distributional measures make that risk governable. They identify dependency risk and help leaders ask whether case assignment, support, schedule, or compensation design is creating an avoidable imbalance. They are not a judgment about the people carrying the load.

5.1 Systems theory and emergence

Capacity is an emergent property of interactions rather than a sum of resources. The relevant unit is the care pathway, not a department silo. Delayed access can increase acuity, backlogs increase interruption, fatigue increases rework, and turnover intensifies workload. Boundaries matter too: an intervention can look effective when costs are shifted to patients, referring clinicians, another modality, or another shift.

5.2 Theory of constraints

A disciplined sequence: identify the constraint, exploit or protect it, subordinate or align the rest of the system, elevate it when necessary, and repeat because the constraint moves. Rawson et al. (2026) applied this logic to radiology. If MRI technologist availability limits access, faster interpretation will not increase throughput. If protocol ambiguity causes lost scanner time, capital expansion is premature.

5.3 Queueing, variability, and buffers

Utilization near theoretical maximum produces disproportionate waiting when demand and service times vary. Radiology variability is structural: emergency demand, protocol complexity, sedation, contrast reactions, add-ons, equipment failure, and inpatient transport are not fully predictable. Eliminating all slack is not efficient in a stochastic system. Buffers should be intentional, visible, and sized to risk, not treated as evidence of laziness.

5.4 High-reliability organizing

Sensitivity to operations, reluctance to simplify, preoccupation with failure, deference to expertise, and commitment to resilience. Evidence for HRO implementation in health care is mixed and often methodologically weak (Veazie et al., 2022; Fricke et al., 2025), and structural barriers such as time, infrastructure, role clarity, and burnout persist (Evans et al., 2026). CAPACITY uses HRO as observable routines, not as a label or certification.

5.5 Job demands-resources

Exhaustion risk rises when demands exceed resources such as control, support, recovery, usable technology, staffing, and meaning. Shanafelt et al. (2015) found each one-point increase in a leadership score associated with a 3.3 percentage-point reduction in burnout likelihood and a 9.0 percentage-point increase in satisfaction after adjustment. The association supports leadership relevance. It does not prove that a one-point intervention produces those exact effects.

5.6 Sociotechnical systems

Technology value does not reside in the algorithm alone. It emerges from model performance, interface design, workflow placement, case selection, human verification, training, governance, downtime response, and payment together. An AI application that saves time in a controlled reading experiment may add burden when deployed through a slow integration or at a point that is not constraining.

1. Identify the constraint

Combine process mapping, direct observation, staff interviews, queue analysis, capacity-loss minutes, and demand patterns. Distinguish the active constraint from visible symptoms. A long MRI wait may be driven by protocoling, authorization, technologist vacancy, nursing, magnet time, or a schedule template mismatch.

2. Protect and exploit it

Shield the constraint from avoidable work and interruption before adding anything. If the constraint is subspecialty attention, moving administrative tasks off it produces more than an additional shift elsewhere.

3. Align the whole system

Subordinate adjacent processes to the constraint. Scheduling, preparation, transport, protocoling, and staffing patterns should be designed so the limiting resource is never idle and never overloaded.

4. Elevate capacity

Add staffing, capital, hours, or service reconfiguration only when redesign cannot safely meet forecast demand. Process improvement must not be used to deny genuinely needed staffing or equipment.

5. Reassess and rebalance

Test whether relieving the candidate constraint measurably changed the clinical aim and whether downstream resources absorbed the gain. If the constraint improved but the aim did not, another constraint has emerged.

The capacity huddle at the center

The cycle is run by a multidisciplinary huddle rather than a single function, because cross-functional constraints need cross-functional authority. Demand, access, workforce, quality, and financial signals are read together, not in separate meetings.

Decision relevance

Every measure should inform an action, an escalation, or a learning question. A measure nobody would act on is burden.

Balanced causality

Combine structures, processes, outcomes, and balancing measures rather than counting activity alone.

Distribution visibility

Report medians with tails, stratification, and concentration wherever those change the interpretation.

Operational definition

Specify numerator, denominator, exclusions, frequency, owner, and data latency for every measure.

Guardrails before targets

Define safety, equity, and workforce floors before any optimization begins, not after a problem appears.

Minimal burden and validation humility

Prefer a compact trusted scorecard to a large unowned dashboard. Label experimental composites and avoid punitive use.

Floors are non-compensatory

A composite score must never authorize unsafe performance. Local leaders should define floors for critical events, overdue urgent studies, significant quality deterioration, unsustainable overtime, or severe inequity. If a floor is crossed, the scorecard status is red regardless of the index.

Missingness must be visible

Imputing favorable performance when data are absent would reward weak measurement. A pilot should report completeness by domain and suppress the index entirely when minimum data-quality rules are not met.

Why the sequence matters

A cancellation cannot occur in a slot already lost to downtime, and a repeat cannot occur on a study that was never performed. Applying rates in parallel double-counts loss and overstates the recoverable opportunity, which is the same error the report identifies in technology business cases that sum overlapping benefits.

The coupling matters too. The constraint may be an asset, a person, or the link between them. Adding scanner hours without technologists or nursing creates nominal capacity that cannot be converted.

What the waterfall is for

It makes loss mechanisms visible without blaming individuals. Some loss is required for safety and maintenance. Some is random. Some can be reduced through standard work, readiness checks, parallel processing, protocol simplification, or better matching of examination type to slot length.

The output feeds the Assets domain of the scorecard as reliable slot yield: completed diagnostically usable studies divided by planned slots, after documented required maintenance.

9.2 Constraint-hour economics

Constraint-hour economics values the resource that currently limits the clinical aim. If a magnet is limiting, the unit may be completed diagnostic studies per magnet-hour. If neuroradiologist attention is limiting, the unit may be clinically completed cases per neuroradiologist-hour, including consultation and communication. If authorization is limiting, the unit may be cleared orders per specialist-hour.

This clarifies opportunity cost and prevents assigning equal value to time released in non-limiting areas. Leaders should also avoid summing overlapping benefits: time saved cannot simultaneously be valued as reduced labor and as additional volume unless both are operationally realized.

9.3 What a technology business case must include

Acquisition, integration, cybersecurity, validation, training, monitoring, support, downtime, model drift, and decommissioning. Benefits should be modeled in conservative, base, and optimistic scenarios, and the adoption curve should be explicit.

A generalist platform may reduce integration burden compared with multiple narrow tools, but broad capability can also increase governance scope (Dogra et al., 2025). The decision should rest on local workflow fit and verified outcome, not on vendor-reported accuracy alone. Molwitz et al. (2026) found the economic value of radiology AI depends on task complexity, price, performance, and workflow conditions, with limited high-quality evidence available.

Days 1 to 15: charter and clinical aim

Name the executive sponsor, clinical and operational co-leads, analyst, improvement lead, and safety lead. Define one clinical aim in patient terms, such as reducing the 90th-percentile order-to-performed interval for oncology MRI while preserving repeat rate, staff overtime, and access equity. Establish the system boundary, meeting cadence, escalation rule, and time commitment. Document current policies and concurrent initiatives that could affect interpretation.

Days 5 to 30: baseline and data quality

Build a minimum viable data set across all eight domains, but limit each domain to one or two measures. Validate timestamps against the frontline workflow. Reconcile definitions across scheduling, RIS, PACS, human resources, equipment, safety, and finance. Plot at least 8 to 12 weeks of baseline, longer for seasonal services. Record missingness. A technically elegant dashboard built on unstable definitions should not advance to scoring.

Days 15 to 40: constraint diagnosis

Combine process mapping, direct observation, staff interviews, queue analysis, capacity-loss minutes, and demand patterns. Distinguish the active constraint from visible symptoms. Test the candidate constraint by asking whether relieving it should measurably change the clinical aim and whether downstream resources can absorb the gain.

Days 31 to 70: interventions and buffers

Use small, reversible interventions. Protect the constraint from avoidable interruptions, move preparation steps earlier, standardize high-frequency protocols, redesign appointment templates, pool appropriate queues, recover cancellations, adjust staffing to demand peaks, or automate low-risk administrative work. Define expected mechanism and effect size before launch. Preserve buffer capacity for urgent demand and measure whether staff burden moves elsewhere.

Days 31 to 90: weekly learning cycles

Review the scorecard weekly with annotations. Separate implementation failure from theory failure. If an intervention is not used, attend to adoption and workflow fit. If it is used but the constraint does not move, the mechanism may be wrong. If the constraint improves but the clinical aim does not, another constraint may have emerged. Treat qualitative feedback as diagnostic data, particularly when it identifies invisible work or burden transfer.

Days 71 to 90: evaluation and decision

State the baseline, intervention, adherence, effect, uncertainty, balancing measures, equity results, financial results, and unintended consequences. The decision options are stop, adapt, extend, or scale. Scale requires sustained improvement, no material deterioration of guardrails, reproducible standard work, clear ownership, and resources for implementation at the next site. A successful local pilot is not automatically evidence of generalizability.

Embed rather than sustain by effort

Sustainability requires integration into budgeting, capital planning, workforce strategy, quality governance, technology lifecycle management, and clinical service planning. A capacity program that depends on the enthusiasm of a council will decay when the council’s attention moves.

Scale mechanisms, not surface features

A same-day slot strategy that improves outpatient MRI may not transfer to ultrasound, emergency CT, or subspecialty interpretation. The transferable elements are the clinical aim, linked measures, constraint diagnosis, protected buffers, balanced guardrails, and learning cadence. Local teams should adapt the intervention while maintaining fidelity to those core functions.

Construct development first

Define domain constructs through Delphi methods, cognitive interviews, and multi-stakeholder content validation. Assess candidate measures for feasibility, reliability, responsiveness, redundancy, and susceptibility to gaming. Examine factor structure, convergent and discriminant validity, and relationships with external outcomes across diverse practice settings. Domain scores should not be finalized from expert preference alone.

Then causal evaluation

A stepped-wedge cluster design may suit phased governance implementation with feasible randomization order. Controlled interrupted time series strengthens inference where a comparable nonintervention service exists. Difference-in-differences may help when parallel-trend assumptions are plausible. Every design should include pre-specified balancing measures, equity stratification, implementation fidelity, and qualitative process evaluation. Short pre-post comparisons without a counterfactual are learning evidence, not effectiveness evidence.

On the evidence

This is a critical integrative review rather than a systematic review. The search and selection process was not designed to support exhaustive retrieval or meta-analysis. Publication bias, rapidly changing evidence, and differences in national workforce and payment systems limit generalizability. Several operational studies were conducted in single institutions and may not translate to other settings.

On the model

The CAPACITY model is original and unvalidated. Its eight-domain structure may omit relevant constructs or combine constructs that should remain separate. The acronym may improve recall but can also create a false sense of completeness. The index has not been tested for reliability, validity, responsiveness, calibration, or unintended behavior. Equal weighting is arbitrary. The geometric mean is theoretically motivated but may be overly sensitive to measurement error in one domain.

On causal claims

Associations between workload, burnout, leadership, turnover, and error do not establish that every proposed leadership intervention will improve those outcomes. The capacity-loss cascade is a plausible mechanism, not a directly estimated structural model. Economic values from AI, reconstruction, workflow, and retention studies depend on assumptions and local context.

On scope

The model is primarily oriented to diagnostic radiology in the United States. Interventional services, pediatric environments, rural access, low-resource settings, and other national systems may require different domains, weights, measures, and governance. Future research should deliberately include diverse settings and patient populations.