Features Derived From Blood Pressure Predict Elevated Intracranial Pressure in Critically Ill Children
Abstract Clinicians frequently observe hemodynamic changes preceding elevated intracranial pressure (ICP) events. We employed a machine learning approach to identify novel and differentially expressed features associated with elevated ICP (eICP) events in children with severe brain injuries. Statistical features from physiologic data streams were derived from non-overlapping 30-minute analysis windows prior to 21 eICP events; 200 records without eICP events were used as controls. Ten Monte Carlo simulations with training/testing splits provided performance benchmarks for 4 machine learning approaches. XGB yielded the best performing predictive models. SHAP analyses demonstrated that a majority of the top 20 contributing features from each simulation consistently derived from blood pressure data streams up to 240 minutes prior to eICP events, rivaling ICP-derived features at 0-60 minutes. Our AUROC benchmark at the 30-60 minutes analysis window using the XGB model bundle was 0.82 (95% CI 0.81-0.83); the AUPRC was 0.24 (95% CI 0.23-0.25), well-above the expected baseline. We conclude that physiomarkers discernable by machine learning are concentrated within blood pressure data up to 4 hours prior to eICP events and demonstrate robust benchmark performance. Future predictive modeling of elevated ICP events should leverage features contained within hemodynamic signals.