Presentation (CTI)

OBJECTIVES: Health technology adoption decisions are made at organizational and system levels, yet supporting evidence is usually patient-level only. Few HTA frameworks unify patient-level heterogeneity with organizational adoption under guideline-constrained implementation. Classical AI methods - probabilistic inference, sequential decision-making, and symbolic reasoning - can be combined to bridge this gap. This study develops and demonstrates a multi-scale decision architecture integrating these methods for system-level HTA. The framework is illustrated through robot-assisted radical prostatectomy in a UK NHS medium-volume centre — a setting where learning curves, capacity reallocation, and guideline-constrained patient selection interact.
METHODS: The architecture operated across three tiers. Tier 1 applied a patient-level Bayesian network over surgical approach, learning-curve phase, complications, length of stay, and episode cost. Tier 2 modelled adoption decisions sequentially via a trust-level Markov decision process over 2,100 procedures across a 14-year horizon at 3.5% discounting. Tier 3 verified trajectories symbolically against 10 NICE NG131 constraints, excluding guideline-inconsistent cases. System-level value, defined as discounted net organizational benefit incorporating cost consequences and monetized capacity gains, was estimated via Monte Carlo simulation. Inputs were drawn from a 33-source systematic review validating 44 parameters, under observational (BAUS 2019) and RCT-anchored configurations.
RESULTS: Symbolic verification rejected 49.6% of simulated trajectories. Though cumulative value declined only 2.5%, guideline-inconsistent pathways concentrate in lower-value regions. Median system-level value was £8.5M (P10-P90: £8.0M-£8.9M); per-procedure value was £3,539 (P10-P90: −£2,915 to £15,213), reflecting variability in case-mix and learning-curve position rather than model imprecision. Length-of-stay reductions dominated aggregate benefit. Learning-curve and guideline effects were modest overall (1.2% relative change) but affect when value accrues.
CONCLUSIONS: Combining these three classical AI methods extends conventional HTA by making organizational learning and adoption dynamics explicit, rather than treating them as fixed background. The approach is most useful where capacity reallocation, learning curves, or guideline-driven patient selection materially shape a technology's real-world value.