Introduction: There is currently no tool to support personalised mean arterial pressure (MAP) targets in sepsis. Clinical practice varies widely; meanwhile, routinely collected data remains largely untapped for informing real-time clinical decisions. STRIVE-MAP is a machine learning-based decision support system that estimates the expected clinical benefit of different MAP levels by comparing ICU patients in real time with similar historical cases. Methods: STRIVE-MAP performs behaviour policy evaluation rather than learning a treatment strategy. It uses deep phenotyping and patient matching to estimate the clinical reward of maintaining specific MAP levels. The reward function integrates in-hospital survival and organ recovery to support precision haemodynamic management. We validated STRIVE-MAP on two datasets: a held-out MIMIC-IV test set (USA) and SICdb (Austria). To estimate treatment effect, we built a Marginal Structural Model with Inverse Probability of Treatment Weighting (MSM-IPTW). This emulates a causal longitudinal analysis comparing outcomes under observed care during periods aligned vs misaligned with STRIVE-MAP targets, adjusting for time-varying confounders. Results: STRIVE-MAP generated individualised MAP targets for 26,606 patients in MIMIC-IV (1.35M hours) and 8,794 patients in SICdb (404K hours). Alignment between AI and clinician MAP (within the same 5 mmHg bin, e.g. 65-70 mmHg) occurred in 16.0% (MIMIC-IV) and 14.3% (SICdb) of observations. In 42-44% of cases, AI suggested a lower MAP; in 42%, a higher one. After weighting, covariate balance improved substantially, with all standardized mean differences (SMDs) reduced to < 0.1, indicating successful adjustment for confounding. Alignment with STRIVE-MAP was associated with lower in-hospital mortality (MSM-IPTW OR 0.87, 95% CI 0.84–0.90, p< 0.001 in MIMIC-IV; 0.87, 0.83–0.92, p< 0.001 in SICdb). Conclusions: Under causal modelling assumptions, following the STRIVE-MAP policy was associated with lower marginal mortality risk compared with observed practice. This novel AI system estimates the expected benefit of personalised treatment targets based on historical patient data, avoiding untested AI recommendations and preserving interpretability and safety.
Festor et al. (Sun,) studied this question.