CNN-based deep learning models demonstrated the feasibility of predicting postoperative cerebrovascular events from intraoperative ABP waveforms in pediatric Moyamoya disease patients, with diastolic runoff dynamics emerging as a potentially relevant physiologic pattern.
Key Findings
Results
CNN-based models outperformed Vision Transformer architectures for classifying postoperative cerebrovascular events from arterial blood pressure waveforms.
Models evaluated included ResNet50, ResNet34, DenseNet121, VGG16, VGG19, and Vision Transformer (ViT) architectures
The highest internal classification performance was observed using raw pulse waveforms with an AUROC of 0.772 (SD = 0.070)
Multiple image transformation methods were tested across all architectures
The optimal model configuration used raw pulse waveforms with three consecutive pulses per image
Results
The optimal deep learning model achieved an AUROC of 0.738 ± 0.011 in an independent temporal validation cohort.
The independent temporal holdout cohort consisted of 79 pediatric patients reserved specifically for validation
Performance was measured at the patient level
The training cohort included 181 pediatric patients (≤18 years) who underwent revascularization surgery for MMD
The validation cohort was temporally separated from the training cohort
Results
Grad-CAM visualization identified the diastolic runoff phase of the arterial waveform as the primary region of interest for model classification.
Gradient-weighted Class Activation Mapping (Grad-CAM) was used for explainability analysis
The diastolic runoff phase was consistently highlighted as relevant to the classification decision
This finding suggests the model focused on a physiologically interpretable waveform feature rather than arbitrary signal characteristics
Diastolic runoff dynamics were described as 'a potentially relevant physiologic pattern'
Results
Four waveform-derived features related to arterial compliance were significantly associated with postoperative cerebrovascular events.
All four features reached statistical significance at p < 0.05
The features were specifically related to arterial compliance characteristics
Statistical comparisons were conducted between patients with and without postoperative cerebrovascular events
These features were derived from the intraoperative ABP waveforms after preprocessing including detrending, pulse segmentation, and normalization
Methods
The study included 181 pediatric patients with Moyamoya disease undergoing revascularization surgery, with postoperative cerebrovascular events including transient ischemic attacks, infarctions, and hemorrhages.
All patients were aged ≤18 years
The study was retrospective in design
Postoperative cerebrovascular events encompassed transient ischemic attacks, infarctions, and hemorrhages
An independent temporal holdout cohort of 79 additional patients was reserved for validation
ABP signals were preprocessed using detrending, pulse segmentation, and normalization, then converted into image representations
Conclusions
The authors characterized their findings as exploratory and requiring prospective multi-center validation before clinical application.
The study was retrospective and single-center in design
The authors explicitly stated findings 'are exploratory and require prospective multi-center validation before clinical application'
Despite validation in a temporal holdout cohort, generalizability to other centers and populations remains unestablished
The study aimed to develop an 'explainable' model, incorporating Grad-CAM analysis alongside predictive performance metrics
What This Means
This research suggests that artificial intelligence can analyze the shape of blood pressure waves recorded during surgery to predict which children with Moyamoya disease — a rare condition causing narrowed brain arteries — are at higher risk of having a stroke or similar brain event after their operation. The researchers trained several types of deep learning models on data from 181 pediatric patients and tested the best-performing model on a separate group of 79 patients, finding that the model could distinguish higher-risk from lower-risk patients with moderate accuracy (a score of about 0.74 out of 1.0 in the validation group). Importantly, the AI focused on a specific part of the blood pressure wave called the 'diastolic runoff phase,' which relates to how blood flows out of the arteries during the relaxation phase of the heartbeat — a finding that makes physiological sense and adds credibility to the model's behavior.
The study also found four specific measurable features of the blood pressure waveform, all related to how stiff or compliant the arteries are, that were statistically linked to whether a child went on to have a postoperative brain event. This suggests that the mechanical properties of a child's arteries, as reflected in routine intraoperative monitoring data, may carry important information about surgical risk that is not currently being systematically captured or used in clinical decision-making.
This research suggests that existing monitoring equipment already in use during surgery could potentially be leveraged with AI tools to provide real-time risk information for this vulnerable patient population. However, the authors themselves caution that these findings are preliminary and exploratory, based on a single-center retrospective study, and that larger prospective studies across multiple hospitals would be needed before this approach could be considered for actual clinical use.
Park J, Shin Y, Kim J, Kim Y, Lee S, Kim E, et al.. (2026). Deep learning-based arterial waveform analysis for predicting postoperative cerebrovascular events in pediatric patients with Moyamoya disease.. PloS one. https://doi.org/10.1371/journal.pone.0350637