An iterative, life-cycle approach to the evaluation of healthcare technologies requires that clinical and economic evidence is collected since the initial stages of diffusion. Nevertheless, early cost-effectiveness models are challenging mainly due to the difficulties in estimating model parameters and faithfully characterizing parameter uncertainty. This is especially true with AI-based diagnostics, where attribution of effects on costs and patient-relevant outcomes is more challenging. Empirical applications of early-models are useful to identify the main challenges of iterative modelling and provide recommendations on best-practices. Here, we reported on a case study on a machine learning-derived hypotension predictive index (HPI), that predicts the onset of intraoperative hypotension and trigger corrective measures.

A hybrid decision-tree/Markov model was developed comparing an HPI-based intervention protocol to standard-of-care intervention protocol during gynecological procedures. A short-term component of the model was populated using data from individual patients collected at one hospital in Italy. An historical control group was also defined using propensity score matching. Long-term costs and consequences of HPI were modelled using secondary data. A probabilistic version of the headroom approach was used to determine the maximum achievable price of HPI based on available evidence. Value of Information analysis was also conducted to identify the parameters that contribute the most to the overall uncertainty, and to identify optimal future study designs. Extensive deterministic and probabilistic sensitivity analyses were conducted to characterize the uncertainty over the cost-effectiveness of HPI.

The preliminary results of the analysis suggest that HPI has potential to improve patients’ outcomes and generate efficiency gains by reducing hypotension events and permanent complications, such as acute kidney injury. The link between reduction in hypotension and the rate of complications, or the long-term effects on healthcare costs and patients’ quality of life are the parameters that contribute the most to model uncertainty.

Early cost-effectiveness models are a valuable tool to inform further product development and evidence requirements, but characterization of uncertainty and transparency in modelling assumptions are key.