Research over the past two centuries has shown that maintaining a balance among intracranial compartment volumes is crucial for cerebral function, as elevated intracranial pressure (ICP) reduces cerebral blood perfusion and disrupts cerebral blood flow regulation¹. The relationship between intracranial volume and pressure, serving as a proxy for intracranial compliance (ICC)², stands as a vital indicator of the state of intracranial dynamics.

In addition to monitoring factors like brain tissue oxygenation, metabolism, and electrical activity, assessing ICC holds paramount importance in the field of neurocritical care³. Nevertheless, the gold standard for ICP monitoring, as well as previously proposed ICC monitoring systems, involve placing an invasive probe into the ventricle or brain parenchyma^4,5. The invasive method, while effective, is limited by the need for specialized personnel to perform the procedure, the substantial cost, and the inherent risks associated with invasive brain procedures.

To address the broader context of elevated ICP, it is essential to recognize the diverse underlying causes contributing to this condition, including traumatic brain injury (TBI), intracranial hemorrhage, hydrocephalus, tumors, and severe ischemic stroke. These conditions disrupt the delicate balance among intracranial compartments, leading to increased ICP and associated complications, such as reduced cerebral perfusion and potential herniation. Identifying populations that could benefit from noninvasive ICP measurement is equally crucial. High-risk groups, such as patients in neurocritical care units, individuals with chronic neurological disorders, or those requiring immediate ICP assessments, such as in the battlefield or emergency settings, would significantly benefit from a safer, noninvasive alternative. The primary advantages of noninvasive methods lie in their potential to reduce procedural risks, enhance accessibility to ICP monitoring in resource-limited settings, and provide dynamic, real-time assessments.

In the pursuit of safer medical practices, modern noninvasive techniques have emerged, offering physicians valuable tools to investigate intracranial hypertension (IH) across various clinical scenarios. These techniques include transcranial Doppler (TCD), optic nerve sheath diameter (ONSD) ultrasonography, pupillometry, etc. Extensive research has been conducted to understand their advantages and limitations, with the consensus indicating satisfactory negative predictive value but low positive predictive power to identify elevated ICP⁶.

TCD was proposed as a method to monitor ICC via a computational approach that takes into account the compartmental compliances of the cerebral arterial bed and the cerebrospinal space. However, direct ICP monitoring could not be replaced with this approach⁷. More recently, an additional noninvasive method has joined this array of tools for investigating ICC and IH—the brain4care (B4C) System⁸. This novel mechanical sensor can detect micrometric pulsatile cranial expansions originated from ICP variations within each cardiac cycle. The B4C System has demonstrated the ability to capture surrogate ICP pulse morphology (ICP waveform—ICPW), a physiological “vital sign” closely associated with ICC⁹. The high sensitivity of the B4C System allows it to detect brain pulsations with amplitudes ranging from 0.04 to 0.80 mm¹⁰, which result in corresponding micrometric pulsatile cranial expansions. The system also offers automated waveform analysis and translates ICP variations into numeric parameters (such as P2/P1 ratio, time-to-peak [TTP]) in real-time at the bedside, enhancing the dynamic monitoring of patients^11,12. The P2 wave (tidal wave) follows the P1 wave (percussion wave) in the ICP waveform. An elevated P2/P1 ratio indicates reduced ICC, suggesting the brain’s diminished ability to compensate for increased intracranial volume. Additionally, the TTP (duration from the start of the ICP wave to its highest peak) provides information about intracranial system responsiveness: shorter intervals indicate high ICC, while longer intervals indicate low, detrimental ICC¹³.

Advancements in artificial intelligence (AI), particularly in the realm of machine learning (ML), have shown remarkable potential for brain monitoring, whether it be for clustering arterial blood pressure (ABP) or TCD waveforms to estimate ICP¹⁴ and ICPW¹⁵ assessments, or even for identifying complications in ICP monitoring, such as ventriculitis¹⁶.

Despite these promising developments, the delivering of noninvasive bedside ICP estimation remains elusive. Thus, the primary objective of the present study is to leverage a substantial dataset of ICPW recordings obtained with the B4C System and employ ML techniques to estimate ICP values noninvasively in patients with concurrent invasive ICP monitoring in the neurocritical care setting. Secondarily, this model is compared with a previously reported noninvasive method based on TCD to provide a baseline assessment of clinical performance.

A total of 136 patients were considered for assessment. Ten patients were excluded from the initial set due to the low quality of pooled noninvasive ICPW (nICPW) morphologies (such as the presence of movement artifacts and unreliable waveforms) or the short duration of monitoring, resulting in a revised pool of 126 patients. Subsequently, an additional 14 patients presenting decompressive craniectomy were removed, resulting in 112 patients whose data were segmented into 10-s windows, resulting in a total of 11,604 windows, equivalent to ⁓150,000 pulses (Fig. 1). Table 1 represents the population sample demographics and data allocation for model development and model validation.

The figure provides an overview of the data source, patient selection procedures, and model training workflow.

Data from 92 patients were preprocessed (9591 10 s windows) and employed to generate and train the ML model. Overall, the entire population sample (n = 112) exhibited 5.36% of data with values above 20 mmHg (Table 1). In the model development sample, 86.6% of patients had an external ventricular drain (EVD), and 14.4% of patients had an intraparenchymal ICP transducer, whereas in the validation sample, 100% had an EVD.

ICP was 11.89 ± 3.21 for windows <20 mmHg (n = 1996) and 21.26 ± 1.67 for windows >20 mmHg (n = 17) in the validation sample. The negative predictive value (NPV) was 0.99 and the positive predictive value (PPV) was 0.14.

The model’s cross-validation folds showed a mean absolute error (MAE) of 1.23 ± 0.02 mmHg and a mean squared error (MSE) of 3.68 ± 0.34 mmHg for the test dataset. For the train dataset, the MAE was 0.91 ± 0.01 mmHg, and the MSE was 1.88 ± 0.05 mmHg. Table 2 shows the model’s cross-validation error distribution for each of the 10 folds. The MAE for the validation dataset was 3.00 mmHg, and the MSE was 13.56 mmHg (n = 20, 2013 10-s windows). These values are synchronous with changes in actual ICP, for example as shown in Fig. 2.

The figure displays an overview of 2013 10 s segments of concurrent mean invasive ICP (blue) and noninvasive estimated ICP (eICP) (red) values obtained from the validation dataset (n = 20). Individual trends per patient can be observed in Supplementary Figs. 1–3.

Table 3 shows the noninvasive estimated ICP (eICP) values for the validation dataset individually, which displays the average values of ICP and eICP, along with the observed difference for each patient. A graphical analysis of ICP and eICP trends in time for each validation patient is presented in Supplementary Figs. 1–3.

The Bland–Altman analysis revealed a mean difference (bias) between ICP and eICP of −0.21 mmHg and an SD of ±3.68 mmHg. From these values, a 95% confidence limit for ICP prediction of less than ±7.5 mmHg was achieved (Fig. 3). There was a moderate relationsh