Background/Aims:
Early systolic blood pressure (SBP) lowering is guideline-recommended in acute intracerebral haemorrhage (ICH), but patients at risk of adverse outcomes remain unidentified. This study aimed to identify personalised SBP trajectories using an outcome-aware clustering approach to inform optimal BP control in acute ICH management.
Methods:
This study pooled individual patient-level data from three INTERACT trials with harmonised BP-lowering protocols (targeting SBP <140 mmHg [intensive] vs. <180 mmHg [guideline] within 1 hour). Optimal SBP trajectories were defined by favourable functional outcome (modified Rankin Scale [mRS] 0–2). Patients were first grouped using an outcome-aware clustering approach, with cluster number selected by a weighted Calinski–Harabasz (CH) index incorporating baseline characteristics, first 24 hours SBP, and functional outcome (mRS). The cluster enriched with favourable outcomes was used to inform subsequent clustering within each baseline SBP range, from which optimal SBP trajectories between 30 minutes and 24 hours were derived for each cluster. For a new patient, baseline SBP and baseline characteristics were used to assign the patient to the most similar cluster, with the corresponding cluster-level SBP trajectory serving as the personalised optimal trajectory.
Results:
8011 ICH patients (mean age 62.5 y, 36.9% female) were included. Initial analysis identified two clusters. Subsequent clustering across 14 baseline SBP bins yielded 28 optimal SBP trajectories. Patients with baseline SBP ≤175 mmHg showed stable, predictable trajectories, whereas higher SBP groups exhibited greater variability and prediction errors.
Conclusion:
A semi-supervised machine learning framework showed potential for identifying cluster-level optimal SBP trajectories for BP management in acute ICH.