Abstract
PURPOSE: Elevated intracranial pressure (ICP) contributes to poor neurological outcomes in brain injured patients but relies on invasive approaches with inherent risks. ICP dynamics are closely related to cerebral blood flow velocity (CBFV) and arterial blood pressure (ABP). Data-driven, non-invasive techniques show promise but lack clinically acceptable accuracy and require large amounts of data. METHODS: We used a pre-trained transformer-based foundation model (MOMENT) to generate embeddings from ABP and CBFV in two cohorts a) Columbia University Irving Medical Center (CUIMC) b) Medical Information Mart for Intensive Care III (MIMIC). These embeddings were then used to train a stochastic gradient descent regressor for deriving non-invasive ICP (nICP). Model performance was evaluated using median absolute error (MAE) and Bland-Altman statistics, with leave-one-patient-out validation strategy for within-cohort testing. Further models trained in one cohort were validated on the other to assess generalizability. RESULTS. The CUIMC dataset included 11 patients with a median age of 56 years [IQR: 44–66], and ICP ranging from 0 to 39.8 mmHg. The MIMIC dataset included 11 patients, with a median age of 11 years [IQR: 6.7–17], and ICP ranging from 0 to 35.6 mmHg. The CUIMC-trained model achieved a median MAE of 3.4 mmHg [IQR: 2.54–5.69] on CUIMC dataset and 3.21 mmHg [IQR: 2.46–3.81] on MIMIC dataset. The MIMIC-trained model achieved a median MAE of 3.99 mmHg [IQR: 2.56–5.96] MIMIC dataset and 1.97 mmHg [IQR: 1.87–3.82] on CUIMC dataset. CONCLUSIONS. We present a framework for leveraging a time-series foundation model for deriving robust nICP measures across diverse patient populations.