In Reply: Predicting Intracranial Pressure Levels: A Deep Learning Approach Using Computed Tomography Brain Scans

To the Editor: We thank Palmiero et al for their Letter to the Editor1 regarding our article2 and appreciate the opportunity to clarify the rationale behind our key methodological choices. Concerning the study design and patient selection, our work was conducted as a retrospective study covering a 12-year period (2012-2023). All patients included had both a brain computed tomography (CT) scan and an invasive intracranial pressure (ICP) measurement within ±3 hours in most cases, and within ±1 day when clinically justified. Because of the retrospective design, the decision to insert an ICP monitoring catheter was not part of our study protocol but was made in real time by the treating teams in the emergency department and intensive care unit. These decisions followed the presenting neurological status, CT findings, and established recommendations such as the Brain Trauma Foundation criteria for severe traumatic brain injury3 and broader intensive care unit guidelines for other acute neurosurgical conditions.4 We deliberately did not restrict our data set to severe traumatic brain injury cases alone. The goal was to develop and test a model in a cohort reflecting the mixed pathology encountered in real neurosurgical emergency care (traumatic and nontraumatic) including intracranial hemorrhage, tumors, hydrocephalus, and postoperative states. This broader inclusion enhances ecological validity and challenges the model with the variability seen in daily practice. Concerning imaging markers and rationale for exclusion, classical radiological signs of raised ICP (such as compression of the basal cisterns, midline shift, or diffuse parenchymal swelling) were not used as inclusion or exclusion criteria. This decision was made for two main reasons: First, in a mixed-pathology cohort, these markers are not always reliable indicators of actual ICP status. Mass lesions, chronic changes, or surgical alterations can obscure, mimic, or exaggerate such findings.5 For example, a tumor may cause midline shift without elevated ICP, whereas acute hydrocephalus can raise ICP without marked cisternal compression. Second, preselecting only cases with overt markers would bias the data set toward severe or late-stage presentations. By training without such restrictive filters, we allowed the model to learn from the full spectrum of presentations. This also enables the explainability phase, using Class Activation Maps, to evaluate whether the network autonomously recognizes these established radiological correlates and how consistently it applies them.6 Regarding imaging quality and preprocessing, we agree that standardized, high-quality imaging is essential for AI in medical applications. All CT scans underwent a uniform preprocessing pipeline including Hounsfield unit conversion, windowing, resampling to consistent voxel dimensions, normalization, and automated skull segmentation, to harmonize data from different acquisition periods and minimize noise. While a small number of scans (eg, Figure 10 in our paper) contained artifacts, we intentionally retained them to improve robustness to real-world Picture Archiving and Communication System (PACS) variability, as recommended for clinically deployable AI.6 With respect to data set size and generalization, although deep learning models benefit from larger data sets, the rarity of paired ICP-CT data makes such collections challenging. Our data set of 578 cases is among the largest reported for this specific task.7 To mitigate overfitting, we applied transfer learning from pretrained 3D architectures, targeted data augmentation, and class balancing,8 all methods shown to support clinically relevant performance in smaller data sets. As noted in our Discussion, this study was designed as a proof-of-concept, and a multicenter, guideline-aligned expansion is already in progress. Concerning clinical integration of AI, we fully agree that AI should complement, not replace, expert neurosurgical judgment. Recent work using recurrent machine learning models to predict intracranial hypertension in neurointensive care patients confirms both the potential and the current limitations of such tools, and underscores the importance of robust, multicenter data sets for generalization.9 Our vision is integration into PACS as a triage aid, flagging high-risk cases for expedited review and potentially reducing time to intervention (while leaving final decision) making in the hands of the treating clinicians. Finally, while reviewing the Letter to the Editor,1 we noted that one of the references cited (ref. 5) could not be retrieved in PubMed or any other indexing service in the form provided because it conflates title, authors, DOI, and page numbers from unrelated publications. In addition, another (ref. 4) appears to correspond to a more recent and fully indexed publication in the Journal of Neurosurgery.10 These observations do not alter the main points of their critique but are mentioned for completeness. We appreciate the constructive dialogue and the shared goal of advancing safe, evidence-based AI integration into neurosurgical care. By training on a broad, mixed- pathology cohort without restrictive imaging filters, our study not only explored predictive performance but also laid the groundwork for explainability analyses and multicenter validation. We believe this is a realistic step toward noninvasive ICP estimation using widely available CT data.