PaperlayerView exact revision on arXiv

Source-linked AI summary

FLARE: A Systematic, Uncertainty-Aware Framework for Evidence-Based Adoption of Artificial Intelligence in Healthcare

Jacob Idoko, Siddhartha Paudel, Mariana Bento, Roberto Souza, Gouri Ginde

arXiv:2608.23643v1cs.AI

TL;DR

Healthcare AI adoption lacks sufficient evidence about economic viability and full financial and operational implications beyond model performance. FLARE combines fuzzy logic, time-driven activity-based costing, and ROI analysis for lifecycle evaluation, demonstrated in AI-assisted stroke imaging. The case study identifies an approximately 3,992-patient annual break-even threshold and positive first-year ROI at about 5,000 annual stroke patients under expected assumptions.

  • Problem

    Healthcare AI evaluations often emphasize model accuracy while providing limited evidence about economic viability and comprehensive adoption costs.

  • Method

    FLARE combines fuzzy logic, time-driven activity-based costing, and ROI analysis to evaluate AI costs and value across the adoption lifecycle under uncertainty.

  • Results

    The framework was demonstrated through an early health technology assessment of AI-assisted large vessel occlusion detection in the CT stroke pathway for acute ischemic stroke.

  • Takeaways & Limitations

    FLARE provides a transparent framework for early-stage, evidence-based evaluation of the economic implications of AI adoption in healthcare.

  • Takeaways & Limitations

    TDABC time estimates may rely on average values that do not capture the full distribution of activity durations and resource use.

Abstract

from arXiv · show

Artificial intelligence is increasingly being introduced into healthcare workflows, yet most evaluations emphasize model accuracy rather than whether adoption is economically worthwhile in real clinical settings. This study proposes FLARE, a systematic and uncertainty-aware framework for evaluating the financial and operational implications of adopting AI in healthcare. FLARE combines fuzzy logic, time-driven activity-based costing, and return on investment analysis to estimate the cost of clinical service delivery, the cost of AI development and operation, and the economic consequences of workflow integration under uncertainty. The framework was demonstrated through an early health technology assessment case study of AI-assisted large vessel occlusion detection in the CT stroke pathway for acute ischemic stroke. The case study shows how FLARE can quantify conventional pathway cost, AI-related development and recurring costs, and AI-enabled service savings within a unified activity-based model. Under expected assumptions, the analysis identified a break-even threshold of approximately 3,992 patients per year, with positive first-year return on investment at typical annual stroke volumes of about 5,000 patients. The results further show that economic benefit depends not only on algorithmic performance, but also on patient volume, verification time, infrastructure choices, and workflow design. FLARE provides a transparent and practical decision-support framework for early-stage evaluation of AI adoption in healthcare. By making uncertainty, resource use, and implementation trade-offs explicit, it helps clinicians, administrators, and policymakers determine when AI deployment is economically viable and where operational changes may improve value.

1 Introduction

Healthcare AI evaluations often emphasize model accuracy while providing limited evidence about economic viability and full workflow costs. FLARE addresses this gap by combining uncertainty-aware activity-based costing with lifecycle ROI analysis and demonstrating it in an AI-assisted stroke pathway.

  • FLARE estimates the ROI of adopting AI across development, validation, deployment, integration, and sustained healthcare use.
  • The framework combines fuzzy logic, time-driven activity-based costing, and ROI analysis to quantify AI adoption’s economic value under uncertainty.
  • FLARE extends costing beyond clinical service delivery to include one-time AI development or acquisition costs, recurring costs, and project-specific costs.
  • It produces annual and cumulative ROI, net savings, and break-even measures from fuzzy time and cost estimates across expected, optimistic, and pessimistic scenarios.
  • The framework is demonstrated through an early health technology assessment of AI-assisted large vessel occlusion detection in the CT stroke pathway.

2 Background and Related Work

Prior work identifies limited economic evidence for healthcare AI and concerns that conventional costing methods inadequately represent resource use and uncertainty. The background motivates combining TDABC with fuzzy logic for transparent, activity-level evaluation.

  • Existing AI evaluations often emphasize algorithmic performance while offering limited evidence about clinical value, workflow improvements, time savings, or cost reductions.
  • ROI claims may omit organizational investments in integration, infrastructure, data preparation, clinician evaluation, and continuous monitoring.
  • Time-driven activity-based costing: TDABC estimates healthcare costs through detailed process mapping, resource-level costing, and the activity-level formula Activity Cost = Time to Perform Activity × Capacity Cost Rate (CCR).
  • Time-driven activity-based costing: TDABC uses capacity cost rate and activity time as core parameters, with capacity cost rate calculated from resource-group cost divided by practical capacity.
  • Limitations and fuzzy logic: TDABC still relies on time estimates that may not capture variation across patients, conditions, interruptions, and human judgment.
  • Limitations and fuzzy logic: Fuzzy logic complements TDABC by representing imprecise information and incorporating approximate minimum, most-likely, and maximum values for uncertain healthcare activities.

3 Methods

FLARE uses FL-TDABC as an uncertainty-aware costing backbone, combining fuzzy activity times, resource capacity rates, and lifecycle cost modeling for AI-integrated healthcare workflows.

  • FL-TDABC Service Costing: The seven-step process defines the service, maps the value chain and activities, links resources, estimates times and resource costs, computes CCRs, and calculates total fuzzy cost.The framework retains Kaplan and Porter’s seven foundational steps while adding fuzzy time representation.
  • FL-TDABC Service Costing: FL-TDABC replaces deterministic activity durations with fuzzy time estimates to represent variability in healthcare workflows.Triangular fuzzy estimates capture minimum, most likely, and maximum times, allowing uncertainty to propagate through costing.
  • FL-TDABC Service Costing: CCR_i = Annual Cost of Resource i / Practical Capacity of Resource i converts annual resource costs into costs per unit of productive time.Practical capacity excludes non-productive time such as breaks, meetings, and leave.
  • FL-TDABC Service Costing: Total service cost multiplies activity time by the corresponding resource CCR, producing expected, lower, and upper cost bounds.The bounds correspond to optimistic and pessimistic scenarios and can be scaled from one service cycle to annual volume.
  • FLARE Integration: FLARE combines transparent TDABC costing with fuzzy logic to support ROI modeling across AI development, validation, deployment, integration, and sustained use.The framework is intended to structure resource implications and ROI for early AI-adoption decisions.
  • AI Lifecycle Costing: For AI development, FL-TDABC models measurable human and non-human activities and scales per-sample costs across cohorts or datasets.Development costs are categorized as one-time development or acquisition, recurring, and project-specific costs.

4 Results

The case study defines the conventional CT stroke pathway from ED triage through imaging, interpretation, and possible EVT, then estimates its activity costs using FL-TDABC.

  • Cost Estimation: FL-TDABC combines the observed activity structure, fuzzy times, personnel rates, and material costs to estimate the complete CT pathway cost per patient.The detailed process map and costing inputs are provided separately in the paper’s methodological appendix.
  • CT Stroke Pathway: The CT stroke pathway begins with ED triage, continues through CT/CTA imaging and interpretation, and may end with EVT in the NIR suite.The activity structure was adapted from prior direct-to-angiography and CT-based ischemic-stroke workflows.
  • Time and Cost Inputs: The prior workflow mapping lacked activity-level time estimates, so a radiologist time study supplied minimum, most likely, and maximum durations.Personnel rates were sourced primarily from the Job Bank of Canada, with ALIS Alberta or Glassdoor used when needed.

Appendix A.

The appendix applies FL-TDABC to estimate conventional pathway costs, AI lifecycle costs, and AI-assisted service costs for LVO detection in acute ischemic stroke.

  • Conventional Pathway Cost: The conventional CT pathway costs 12099.45 CAD expected, 11491.47 CAD low, and 13011.22 CAD high per patient.These values are obtained by summing fuzzy activity costs across the complete pathway.
  • Conventional Pathway Cost: At approximately 5,000 annual strokes, the estimated provincial CT pathway cost is 65,056,100 CAD.The annual estimate scales the per-patient pathway cost by annual stroke volume.
  • AI Development Costs: AI development costing uses cohort-specific FL-TDABC estimates based on reported data, literature references, expert estimates, and explicit assumptions.The analysis focuses on Heidelberg training and internal test cohorts with sufficient timing detail and excludes unavailable inputs from subtotals when necessary.
  • AI Development Costs: The one-time development estimate represents in-house human-resource expenditure for model development and validation, excluding compute and infrastructure charges.Infrastructure was locally absorbed in the source study, so those charges are not included in this total.
  • Recurring Costs: Recurring operating costs include internally managed model upkeep and AWS SageMaker deployment, with infrastructure modeled in Canada–Central because GPU-equipped instances were unavailable in Calgary.The model assumes one ML engineer at 44.10 CAD/hour and uses AWS pricing estimates for infrastructure.
  • AI-Integrated Service: AI integration changes only CTA interpretation: automated LVO analysis replaces manual interpretation while expert verification remains unchanged.AI inference cost is accounted for in AI development and recurring components rather than charged again within service cost.

Savings and Benefits.

The savings analysis separates service-level benefits from AI costs and distinguishes first-year investment from later recurring-cost periods to evaluate net savings and ROI over time.

  • Service Savings: 54.65 CAD less per patient is the average cost of the AI-assisted pathway compared with the conventional pathway.The service-level saving excludes AI development and recurring costs and scales with annual patient volume.
  • Service Savings: Annual benefit scales per-patient savings by the hospital’s total annual stroke-patient volume.If volume and service times remain stable, annual benefit is similar in later years.
  • ROI Accounting: Year 1 includes both one-time development and recurring AI costs, whereas Year 2 onward includes only recurring costs.Consequently, the overall cost burden is highest in the first year and stabilizes thereafter.
  • ROI Accounting: Net savings are lower in Year 1, while the deficit narrows from Year 2 onward as only recurring costs remain.The analysis uses this distinction to track the financial trajectory after initial development investment.
  • Break-even Analysis: The break-even analysis identifies when cumulative AI-generated savings equal total AI costs.Under pessimistic assumptions, break-even occurs later because workflow gains are smaller and verification time remains higher.

Break-even Analysis.

FLARE estimates break-even by comparing AI system costs with expected per-patient savings, separating total investment recovery from development-only recovery. Under expected assumptions, break-even occurs at approximately 3,992 patients, or 9.58 months at 5,000 patients annually.

  • Total investment break-even: 3,992.05 patients marks expected break-even when total AI system cost is divided by expected per-patient savings.The calculation includes expected one-time development and annual recurring costs.
  • Total investment break-even: 9.58 months is the expected time to break even at an annual volume of 5,000 stroke cases.At this point, cumulative savings offset the first-year AI investment.
  • Break-even caveat: Yearly net savings may become negative when recurring costs exceed the annual benefit.This condition limits the interpretation of break-even as a continuing annual gain.
  • Development-only break-even: 545.78 patients represents expected break-even for development cost alone, excluding recurring operational costs.This calculation reflects only service-side savings paying down the development investment.

Return on Investment (ROI).

FLARE reports first-year and subsequent-year ROI using expected yearly costs and savings. The expected calculation yields positive ROI in Year 1, with higher returns after development costs are excluded.

  • Annual ROI: Roughly 45 cents of net profit per dollar is expected in subsequent years as development costs are excluded.The reported increase reflects the removal of one-time development costs from later-year calculations.

Cumulative ROI over Five Years.

Cumulative ROI improves as one-time development costs are amortized, while patient volume and implementation efficiency determine how quickly the AI investment becomes viable. Under expected assumptions, ROI is positive at typical annual stroke volumes and rises with scale.

  • Five-year cumulative ROI: Approximately 44% cumulative ROI is reached by Year 5 in the expected scenario.The expected case shows steady positive returns as one-time development costs are amortized.
  • Five-year cumulative ROI: More than 150% cumulative ROI is reached by Year 5 in the optimistic scenario, while the pessimistic scenario remains negative.The scenarios represent different fuzzy cost and benefit bounds.
  • Patient-volume sensitivity: Approximately 3,992 patients per year is the expected break-even volume, after which ROI continues to rise.At this volume, cumulative savings equal Year 1 cost and ROI becomes positive.
  • Scenario sensitivity: At volumes below approximately 4,000 cases per year, expected annual savings are insufficient to recover the upfront investment.The expected case therefore produces negative net savings and negative ROI at low volumes.
  • Patient-volume sensitivity: 25.25% expected ROI and positive annual net savings occur at 5,000 stroke patients per year.The paper identifies this volume as economically viable under typical operating conditions.
  • Scenario sensitivity: Optimistic conditions permit profitability near or slightly below 3,000 patients annually, whereas pessimistic conditions delay break-even.The pessimistic scenario reflects smaller workflow gains and higher verification time.
  • Patient-volume sensitivity: ROI increases with scale because fixed investment costs are distributed across more patients while per-patient savings accumulate.The time required to recover the initial investment also decreases as patient volume increases.
  • Time to break-even: At 4,000 patients, expected break-even takes approximately 12 months, declining to roughly 8–10 months at 5,000–6,000 patients.The AI solution begins generating net financial benefit within the first year once sufficient volume is reached.

5 Limitations

FLARE’s conclusions are bounded by detailed input requirements, incomplete uncertainty and financial modeling, a single regional case study, and limited external validation. The numerical results require local recalibration and should not be treated as universally applicable.

  • Input and modeling limitations: FLARE requires detailed activity maps, while its results depend on the quality and accuracy of underlying assumptions.The framework uses expert consultation or approximate values when high-resolution inputs are unavailable.
  • User evaluation limitations: The framework has not undergone a structured user study assessing usability, interpretability, or decision-making impact among intended end users.This limitation concerns user-facing evaluation rather than the validity of the underlying economic logic.
  • Input and modeling limitations: Triangular fuzzy numbers model uncertainty in activity durations and cost parameters, but other uncertainty sources may remain unrepresented.The framework may therefore underestimate the full range of real-world uncertainty in some deployments.
  • Input and modeling limitations: FLARE does not formally model inflation, interest rates, or time value of money through discounting.Costs and benefits are expressed in constant monetary terms, although users can adjust parameters over time for scenario analysis.
  • Scope limitations: The case study covers one stroke-imaging workflow in Calgary, Alberta, so numerical results require local recalibration.Personnel rates, material costs, and workflow characteristics vary across institutions and health systems.
  • Validation limitations: FLARE has not been validated against detailed ground-truth costing datasets or compared directly with alternative costing methodologies.The evaluation demonstrates structure, transparency, and uncertainty-aware capabilities rather than accuracy in reproducing true institutional costs.

6 Discussion

FLARE shows that AI adoption economics depend on patient volume, workflow efficiency, and implementation choices, not model performance alone. The framework makes resource use, costs, savings, and uncertainty explicit for evaluating operational strategies.

  • 6 Discussion: FLARE links clinical activities, resource use, time, costs, and savings to show where streamlined workflow and reduced interpretation time generate financial value.The framework supports varying verification time, hosting strategy, staffing, wages, labeling effort, and GPU/server usage.
  • 6 Discussion: Approximately 4,000 patients per year marks the threshold below which annual savings do not offset the Year 1 investment.Below this volume, the AI solution does not become cost-saving in the first year, although this does not establish that it is unsuitable.
  • 6 Discussion: Lower development costs, subscription services, optimized cloud configurations, and shared GPU resources can move break-even to lower patient volumes.These choices reduce initial or recurring expenses within the modeled implementation.
  • 6 Discussion: At approximately 5,000 stroke cases annually, fixed development costs are spread across more cases and per-patient savings accumulate, producing positive first-year returns.The investment is financially viable in hospitals or regions with moderate to high patient throughput.
  • 6 Discussion: Pessimistic assumptions delay cost recovery, whereas improved workflow integration produces positive ROI at lower volumes.The sensitivity analysis therefore links financial outcomes to verification time and workflow efficiency.
  • 6 Discussion: Effective adoption requires workflow redesign, clinician trust, training, and usability alongside algorithmic performance.The discussion identifies clinical integration as a determinant of economic benefit.

7 Conclusion

The paper introduces FLARE as an uncertainty-aware framework for evaluating the financial and operational implications of AI adoption in healthcare. Applied to AI-assisted LVO detection in acute ischemic stroke imaging, it identifies patient volume and workflow integration as central to economic value.

  • 7 Conclusion: FLARE combines fuzzy logic, FL–TDABC costing, and ROI analysis to estimate care-delivery costs and the economic impact of introducing AI.Its activity-based structure is intended to make resource use and implementation trade-offs transparent.
  • 7 Conclusion: The framework was demonstrated through an early health technology assessment of a real-world acute ischemic stroke imaging pathway.The modeled pathway includes ED triage, CT imaging and interpretation, treatment eligibility confirmation, and NIR-suite treatment when indicated.
  • 7 Conclusion: Approximately 3,992 patients per year was the estimated break-even point using the expected defuzzified AI cost.The estimate comes from the acute ischemic stroke imaging case study.
  • 7 Conclusion: At around 5,000 patients per year, the AI solution produced a positive return on investment within the first year of deployment.This result was reported for the modeled stroke pathway.
  • 7 Conclusion: Economic benefit depends on aligning development costs, personnel effort, infrastructure decisions, and clinical workflow integration.Algorithmic performance alone does not guarantee economic benefit.
  • 7 Conclusion: FLARE provides stakeholders with a structured way to assess when AI adoption is economically viable and which operational adjustments may improve financial outcomes.Suggested adjustments include lowering development costs, optimizing verification workflows, and scaling deployment across sites.

C.2 Full Process Map for the Heidelberg Training Cohort

The Heidelberg training-cohort process map treats one data sample as the service unit and assigns cohort activities to personnel and resource rates. Fuzzy activity times are converted into expected and bounded per-sample costs.

  • C.2 Full Process Map for the Heidelberg Training Cohort: One data sample processed through the development workflow is the unit analyzed for the training cohort.The process map distinguishes training activities from test-cohort activities.
  • C.2 Full Process Map for the Heidelberg Training Cohort: Training samples undergo gathering, preprocessing, labeling, inclusion in training, and optionally post-train checks.These activities define the training-cohort value chain.
  • C.2 Full Process Map for the Heidelberg Training Cohort: Each process activity is assigned to a primary role and its resource rate, using triangular fuzzy durations in minutes.The rate-estimation step uses hourly rates and capacity cost rates from the personnel-cost table.
  • C.2 Full Process Map for the Heidelberg Training Cohort: Activity cost is calculated by multiplying expected time by the corresponding capacity cost rate, with minimum and maximum times defining cost bounds.For the Data Gathering example, expected, low-bound, and high-bound costs are 2.68 CAD, 2.01 CAD, and 3.35 CAD.
  • C.2 Full Process Map for the Heidelberg Training Cohort: Summing activity costs yields expected, low, and high per-sample development-cost estimates.The resulting estimates are reported in the training- and internal-test-cohort tables.
Loading 2608.23643v1…