Non-invasive intracranial pressure prediction using machine learning

Invasive intracranial pressure (ICP) monitoring, while highly accurate, is limited in application due to its invasive nature, restricting use to a narrow subset of patients and specific clinical conditions. This limitation underscores the need for non-invasive alternatives to broaden its utility and accessibility. Non-invasive ICP monitoring has been attempted by numerous approaches, including transcranial Doppler ultrasound, near-infrared spectroscopies, and more, but it has not yet reached a clinically acceptable level of maturity, accuracy, and precision. Recently, several groups have attempted to utilize the fluctuations in cerebral blood flow (CBF) due to the cardiac cycle that are measured by diffuse correlation spectroscopy (DCS) and analyzed by machine-learning and model-based methods to estimate ICP
. In this study, we present one of the largest datasets to date, comprising forty-four patients with idiopathic normal pressure hydrocephalus (iNPH), to train a multilevel wavelet decomposition network (mWDN) for non-invasive ICP estimation. Unlike previous approaches, our method relies exclusively on raw CBF signals without the need for feature extraction from pulses or additional variables, such as mean arterial pressure (MAP). This approach maximizes simplicity while improving the model’s robustness and generalizability, enhancing the reliability of non-invasive ICP estimation.

The mWDN model demonstrates significant promise for non-invasive ICP estimation using DCS-derived CBF signals. By capitalizing on wavelet-based feature extraction and a robust training framework, the model achieves clinically meaningful accuracy and effective temporal tracking of ICP fluctuations. Future work will focus on expanding the dataset with additional high-ICP cases, ultimately refining the model’s performance across a broader pressure spectrum and enhancing its clinical applicability.