In neurocritical care, precise monitoring of intracranial pressure (ICP) is crucial for protecting patients who have sustained acute brain injuries, such as traumatic injuries, large strokes, brain tumors, intracranial hemorrhages, and subarachnoid hemorrhages. These conditions often lead to swelling within the confined space of the skull, elevating ICP. Without timely intervention, elevated ICP can result in secondary brain injury, significantly worsening patient outcomes.

Clinicians typically rely on invasive ICP monitors that provide continuous, pulsatile waveform data. Among these devices, External Ventricular Drains (EVDs) are considered the gold standard because they not only measure ICP but also actively treat elevated pressures by allowing cerebrospinal fluid drainage. EVD placement is notably the most frequent neurosurgical procedure in the United States, with over 36,000 procedures performed annually.

Recent advancements in predictive modeling, especially machine learning and deep learning techniques, hold the promise of forecasting dangerous ICP spikes before they occur. This proactive approach could revolutionize patient management. However, despite their central role in ICP management, EVD-derived waveform data has traditionally been excluded from large-scale predictive studies. This exclusion arises because EVDs switch between 'clamped' states—where they accurately record ICP—and 'draining' states, which complicate waveform interpretation. Retrospective datasets typically lack detailed annotations differentiating these states, posing a substantial barrier to the development of robust predictive algorithms. Developing an effective pre-processing framework therefore requires the monumental task of labeling vast volumes of data. 

To address this critical limitation, our team recently introduced a novel machine learning based algorithm called CICL in a paper published in npj Digital Medicine. CICL is an innovative, semi-supervised machine learning framework specifically engineered to process and accurately label EVD-derived ICP waveforms. By automatically categorizing data segments into 'clamped,' 'draining,' or 'noisy,' CICL enables the large-scale incorporation of EVD waveform data into predictive models. This significantly enhances the sample size and diversity of patient cases available for analysis, including those with complex conditions such as severe traumatic brain injuries, extensive strokes, or hemorrhages.

One of the major challenges in developing CICL was the question of how to label an enormous amount of high frequency (>120Hz) ICP waveform data in order to train the model. From a data science and machine learning perspective, we accomplished this using an innovative combination of waveform segmentation, clustering, and then labelling of homogenous clusters to have expert labelling of a large amount of data for training and testing purposes. This strategy, a sophisticated combination of waveform segmentation, clustering algorithms, and expert-guided labeling, efficiently generates a labeled dataset substantial enough for training and validating predictive models. Importantly, this approach is generalizable and can be adapted to analyze various other types of high-frequency, large-scale time-series data beyond neurophysiology.

Ultimately, CICL represents a significant leap toward achieving real-time, predictive management of elevated ICP. We believe this framework will serve as a valuable blueprint for future research, empowering clinicians and data scientists to develop advanced tools that enhance patient care and outcomes for patients with acute brain injuries.