A transparent, uncertainty-aware costing framework shows AI for CT stroke detection can be economically viable in high-volume centres: modeled break-even is about 3,992 patients per year with positive first-year ROI near typical volumes (~5,000), but value depends strongly on verification time, infrastructure choices and workflow design.

Main Finding

FLARE (Fuzzy-Logic, Time-Driven Activity-Based ROI) is a practical, uncertainty-aware framework that combines fuzzy logic with TDABC and ROI analysis to produce lifecycle economic evaluations of AI adoption in healthcare. In an early health-technology-assessment case study (AI-assisted large vessel occlusion detection in the CT stroke pathway), FLARE identified a break-even threshold of ≈3,992 patients/year and positive first-year ROI at typical annual stroke volumes of ≈5,000 patients. The framework shows that economic viability depends not only on algorithmic accuracy but also on patient volume, verification (human-review) time, infrastructure choices, and workflow design.

Key Points

• Purpose: Move beyond accuracy metrics to assess whether adopting an AI solution is economically worthwhile across its full lifecycle (development → deployment → operation).
• Core idea: Integrate fuzzy logic with TDABC to capture activity-level time uncertainty and propagate that uncertainty into ROI and break-even analyses.
• Components:
  • TDABC core: Activity cost = Time to perform activity × Capacity Cost Rate (CCR).
  • Fuzzy modeling: Represent activity times as triangular fuzzy numbers (a, b, c); expected (defuzzified) time x* = (a + b + c) / 3. Use a and c for optimistic/pessimistic bounds.
  • Lifecycle costing: Classify AI costs into one-time development/acquisition, recurring operational costs, and project-specific costs.
  • Economic outputs: Annual and cumulative ROI, net savings, break-even points, and sensitivity analyses under uncertainty.
• Demonstration: Early HTA for CT-based LVO detection in acute ischemic stroke pathway; unified activity-based model quantified conventional pathway costs, AI development/recurring costs, and AI-enabled savings.
• Sensitivity: ROI is sensitive to patient volume, clinician verification time, infrastructure (on-prem vs cloud), and workflow integration choices; algorithmic performance matters but is only one piece of the value chain.
• Transparency: Produces expected, optimistic, and pessimistic scenarios from the same model, improving decision-maker confidence and revealing which parameters drive economic outcomes.

Data & Methods

• Costing methodology:
  • Time-Driven Activity-Based Costing (TDABC) for granular, activity-level resource tracing.
  • Capacity Cost Rate (CCR) = total cost of a resource group / practical capacity (accounts for non-productive time).
  • Activity cost formula: Activity Cost = Time × CCR.
• Uncertainty modeling:
  • Triangular Fuzzy Numbers (TFNs) for activity durations: (a = min, b = most likely, c = max).
  • Defuzzified expected time x* = (a + b + c) / 3; use a and c for lower/upper bounds.
  • Propagate fuzziness through cost and ROI calculations to produce expected, optimistic, and pessimistic outputs without rerunning point-estimate scenarios.
• ROI metrics:
  • ROI = (Total Benefits − Total Costs) / Total Costs × 100%.
  • Produce annual and cumulative ROI, net savings, and break-even patient-volume thresholds.
• Lifecycle cost accounting:
  • Explicitly model one-time development/acquisition costs, recurring operational costs (inference, monitoring, verification), and integration/IT infrastructure.
• Case study specifics:
  • Context: CT stroke pathway for acute ischemic stroke and AI-assisted LVO detection.
  • Outputs: Estimated break-even ≈3,992 patients/year; typical stroke volumes (~5,000) yield positive first-year ROI under baseline assumptions.
  • Sensitivity analyses explored patient volumes, verification times, infrastructure trade-offs, and algorithm performance.
• Implementation note: FLARE follows Kaplan & Porter’s seven TDABC steps but replaces deterministic time inputs with fuzzy estimates (FL–TDABC).

Implications for AI Economics

• Lifecycle perspective is essential: Economic evaluations must go beyond model purchase/deployment cost and include development, integration, clinician verification, monitoring, and infrastructure.
• Move from accuracy-centric claims to operational economics: Algorithm performance alone does not determine value; workflow design, staff time, and volume drive ROI.
• Volume thresholds matter: Hospitals with low patient volumes may not reach break-even; procurement and deployment decisions should incorporate expected case volumes and scaling plans.
• Human-in-the-loop costs are crucial: Verification and review time can materially change ROI; design choices that reduce verification burden (or safely automate more) improve economics.
• Vendor ROI claims are often incomplete: FLARE highlights common hidden costs (integration, IT, continuous validation). Buyers should demand lifecycle cost transparency.
• Decision support and policy:
  • Hospital administrators can use FLARE to assess whether and how to adopt AI tools, choose infrastructure options, and prioritize workflow redesigns that unlock value.
  • Policymakers and payers can use FLARE outputs (break-even volumes, sensitivity to verification time) to design reimbursement or scale-up incentives.
  • Researchers and vendors should collect activity-level time and resource data to improve model inputs and cross-site generalizability.
• Methodological next steps: FLARE is suitable for early HTA and budgeting; as outcome data accrue, it can be extended to NPV/IRR or formal cost-effectiveness/cost-utility analyses that incorporate clinical outcomes and discounting.
• Limitations to note: Results depend on input quality (time estimates, cost rates, and TFN bounds). FLARE’s ROI focus is a practical early filter but does not replace full health-economic evaluations when long-term outcomes and utilities are available.

Overall, FLARE provides a transparent, reproducible, and uncertainty-aware toolkit to inform early-stage economic decisions about AI adoption in healthcare, shifting emphasis from isolated model metrics to operational and financial realities that determine real-world value.