ICP-SWAn: A selective waveform analysis pipeline for conformal subpeak detection in intracranial pressure signals

Intensive care unit management of patients with traumatic brain injury or aneurysmal hemorrhage often requires intracranial pressure (ICP) monitoring. Although the primary goal is to prevent adverse intracranial hypertensive events, mechanical properties of the cerebrospinal system can also be derived from a closer analysis of the ICP signal. The morphology of the ICP signal can be influenced by a variety of physiological components and is therefore highly variable. In particular, on the time scale of a heartbeat, characteristic subpeaks called P1, P2, and P3 are visible most of the time and their analysis may provide valuable insights about cerebrospinal pressure–volume compensation. However, their automatic detection in real time is a challenging task that requires displaying no result in situations of uncertainty to avoid misleading the clinicians. Therefore, we applied several results from the “Learn-then-Test” theoretical framework to our deep learning-based delineation algorithm to identify cases where it is appropriate to refrain from providing a result. The subpeak detection task was performed using recurrent neural networks trained on a 59-patient dataset. Performance was evaluated on two independent test datasets. The first one consisted of 13,086 pulses extracted from 49 recordings and labeled by 3 independent operators. The second one, which is publicly available, consisted of 11,172 pulses extracted from 63 recordings. The mean absolute error on the calculated P2/P1 ratio was 0.061 for the first test dataset and 0.016 for the second test dataset. • In an online setting, P2/P1 ratio calculation should include proper quality control. • Certain precautions must be taken when performing real-time ICP signal analysis. • The benefits of conformal risk control are mostly seeable in a noisy environment.