Non-invasive intracranial pressure waveform reconstruction with deep learning

Continuous intracranial pressure (ICP) monitoring requires invasive instrumentation and reaches only a subset of the critically ill patients who might need it. We tested whether deep learning models trained on routinely acquired extracranial signals can reconstruct continuous ICP waveforms at clinically relevant accuracy. Using data from adults admitted to the intensive care unit at a single quaternary health system, five deep learning architectures were trained on high-frequency arterial blood pressure, photoplethysmography, and electrocardiography waveforms, using invasive intraparenchymal ICP as ground truth; two fusion strategies and three training objectives were evaluated, and models were externally validated on a held-out independent dataset (the MIMIC-III Waveform Database). Performance was assessed by mean absolute error (MAE) and waveform similarity by Pearson correlation (r). Across 158 critically ill adults (∼5,322 hours) from two institutions (Johns Hopkins Hospital, Baltimore; Beth Israel Deaconess Medical Center, Boston), external-validation MAE ranged from 4.276 to 4.946 mmHg and Pearson r from 0.599 to 0.722, with the multiscale encoder-decoder model showing the most favorable MAE–correlation tradeoff. These findings provide the first demonstration that continuous ICP waveform reconstruction from bedside signals generalizes across institutions at clinically relevant accuracy, establishing a foundation for non-invasive ICP monitoring and motivating validation across broader populations and ICP ranges.