Crossover effect: causal machine learning reveals opposing mortality responses to mean arterial pressure targets...

To quantify the heterogeneity of treatment effects of higher versus standard MAP targets on 28-day mortality, to identify distinct clinical phenotypes that respond differentially to MAP targets, and to derive simple bedside decision rules to guide individualised haemodynamic management.

Current guidelines recommend a uniform mean arterial pressure (MAP) target of ≥65 mmHg for septic shock. However, trials like SEPSISPAM found no mortality benefit from higher targets in hypertensive patients. We hypothesised that this null finding results from the masking of heterogeneous treatment effects (HTEs) among phenotypically distinct individuals.

We conducted a retrospective cohort study of 7665 hypertensive patients with septic shock from the Medical Information Mart for Intensive Care IV database (version 3.1). The primary exposure was time-weighted average MAP (TWA-MAP) during the first 24 h, dichotomised into high target (>75 mmHg) versus standard target (65-75 mmHg). The primary outcome was 28-day mortality. Individual treatment effects (ITEs) were estimated using a T-Learner causal machine learning framework with gradient boosting. Patients were stratified by ITE quartiles, and a Classification and Regression Tree (CART) was used to derive bedside clinical rules.

We observed marked HTE (ITE range: -0.54 to +0.86). In the Strong Benefit group (Decompensated Phenotype) - characterised by tachycardia, hyperlactatemia and renal dysfunction - the high MAP target dramatically reduced 28-day mortality from 62.6% to 22.9% (absolute risk reduction = 39.7%, P < 0.001). Conversely, in the Harm group (Frail Phenotype) - characterised by older age and lower heart rate - the high target increased mortality from 14.7% to 33.9% (ARR = -19.2%, P < 0.001). The CART model identified heart rate >105 bpm and creatinine >1.65 mg/dL as simple bedside criteria to identify patients likely to benefit from higher targets.

The impact of higher MAP targets in hypertensive septic shock exhibits a distinct crossover pattern: life-saving in decompensated patients but harmful in frail patients. This HTE explains the null results of previous landmark trials. Utilising simple parameters like heart rate and creatinine can guide a shift from 'one-size-fits-all' targets towards precision haemodynamic management.