A Real-Time Deep Learning Approach for Inferring Intracranial Pressure from Routinely Measured Extracranial Waveforms in the Intensive Care Unit

Abstract

Objective.Intracranial pressure (ICP) is a vital waveform used to assess the conditions of patients with intracranial pathologies, such as traumatic brain injury. The current standard for monitoring ICP requires a catheter to be inserted into the brain, which is extremely invasive and increases the risk of hemorrhage and infection. We hypothesize that ICP can be inferred from extracranial waveforms routinely measured in the Intensive Care Unit (ICU), such as invasive arterial blood pressure (ABP), photoplethysmography (PPG), and electrocardiography (ECG).Methods.We extracted 600 hours of simultaneous ABP, EKG, PPG, and ICP waveforms (125 Hz) across 10 different patients from the MIMIC III Waveform Database. These concurrent recordings were broken into 10-second windows to train on long-short term memory (LSTM) and temporal convolutional network (TCN) models to predict ICP using ABP, EKG, and PPG as input features. These models were evaluated in both a single-patient analysis and multi-patient analysis study.Results.In the single-patient analysis study, TCN's median mean average error (MAE) test loss was 1.60 mmHg in comparison to a median MAE test loss of 1.88 mmHg for LSTM. TCN was used in the multi-patient analysis study, reporting a test MAE loss of 5.44 mmHg.Conclusions.These novel results indicate that it is possible to predict ICP waveforms from extracranial waveforms in a real-time, continuous, and non-invasive manner. This approach could potentially eliminate the risks associated with invasive monitoring and allow for more timely treatment of patients with intracranial pathologies with further follow-up studies.

Derivation, External Validation and Clinical Implications of a deep learning approach for intracranial pressure estimation using non-cranial waveform measurements

“…Recognizing these limitations, recent research seeks to correlate physiological data with neurological conditions 12 . Yet, such studies are often restricted by their sample size, stringent data filters limiting real world application, and lack of external validation limiting generalizability [12][13][14][15] .…”