Seongho Park

Diffusion-FLAIR mismatch (DFM), defined as the discrepancy between diffusion-weighted imaging (DWI) and FLAIR sequences, is a key imaging biomarker used to identify patients with unclear symptom onset or those likely to benefit from recanalization therapy. However, FLAIR imaging is often limited in acute clinical settings due to time constraints, patient instability, or scanner availability. Therefore, alternative approaches to assess DFM without relying on FLAIR are needed. This study aimed to develop a deep learning model that predicts DFM using only B1000 and apparent diffusion coefficient (ADC) images, and to evaluate its performance. This study was conducted using multicenter stroke registry-based cohorts. After excluding cases not suitable for analysis, 2,369 cases were included in the derivation cohort for model development, and 679 cases from two independent stroke centers were included as external validation cohorts. Model performance was assessed based on two binary classification schemes. The Broad classification included all non-match cases, comprising both clearly defined mismatches and ambiguous cases with subtle or focal FLAIR changes, while the Focused classification excluded such ambiguous cases, including only clearly defined match and mismatch instances. In the Focused classification setting, the AI model achieved an AUROC of 0.92 (95% CI: 0.88–0.95) on the external validation dataset, significantly outperforming the average AUROC of human experts, which was 0.82 (95% CI: 0.77–0.87) ( p < 0.001, AUROC difference 0.10). Overall, the Focused classification yielded higher performance than the Broad classification, and this trend was consistent for both the AI model and human readers. We developed a deep learning-based classifier capable of predicting DFM using only B1000 and ADC images, without requiring FLAIR. The model outperformed human experts and demonstrates potential as a complementary diagnostic tool in acute stroke settings where FLAIR imaging is unavailable, supporting its future integration into decision-support systems.

We developed and externally validated a deep learning model to automatically detect new ischemic lesions on serial FLAIR MRI scans in patients with stroke. Manual interpretation of follow-up imaging is labor-intensive and variable, and silent brain infarctions (SBIs) are frequently missed despite their prognostic importance. Using 25,451 paired slices from 1055 patients across two hospitals, we trained a convolutional neural network with supervised contrastive learning to classify new lesion occurrence. The model achieved an area under the receiver operating characteristic curve of 0.89 in both internal and external validation cohorts. To evaluate clinical relevance, we further analyzed an independent asymptomatic cohort of 307 patients with a median follow-up of two years. Patients classified as SBI-positive by the model showed a significantly higher risk of subsequent symptomatic stroke than those without SBI. In multivariable Cox regression adjusted for age and major vascular risk factors, model-positive patients had a 3.8-fold increased risk of stroke recurrence. These findings indicate that AI can identify clinically meaningful SBIs that are under-recognized in routine practice and independently associated with stroke recurrence. Automated lesion detection may provide a reproducible imaging biomarker for risk stratification, supporting standardized interpretation of follow-up MRI and informing secondary stroke prevention strategies.

SARS-CoV-2 is a global pandemic with high mortality and prevalence. Although the disease severity has weakened, it remains a significant global health burden as an endemic disease. Several reports have explored the relationship between intracranial artery dissection and SARS-CoV-2 infection; however, an association with SARS-CoV-2 vaccination has not been established. Here, we describe three cases of vertebral arterial dissection with acute ischemic stroke after SARS-CoV-2 vaccination. This report aimed to raise the clinical awareness of this potential adverse effect, which may facilitate earlier diagnosis and intervention.

Post-stroke Anxiety (PSA) affects 20–30% of stroke survivors and significantly impacts their quality of life (QoL) and rehabilitation outcomes. Traditional emotion assessment methods rely on subjective self-reports, which have a limited ability to capture real-time fluctuations in emotional states. This study proposes a self-supervised learning (SSL) framework combined with a transformer-based emotion classification model to enable continuous anxiety tracking in patients with stroke using smartwatch-derived photoplethysmography (PPG) signals. The SSL model was pretrained on the VitalDB dataset using R-peak-to-peak intervals (RRIs) extracted from electrocardiography (ECG) signals. The pretrained encoder is then integrated into a transformer-based classifier trained on the Psycho-physiology of Positive and Negative Emotions POPANE dataset containing labeled emotional responses. Finally, the trained model was applied to smartwatch-derived pulse peak intervals (PPI) from patients with stroke, and the predictions were applied to Generalized Anxiety Disorder-7 (GAD-7) scores and we evaluated it against GAD-7 scores using both group-level and 90-day longitudinal analyses, alongside on-device feasibility on a Galaxy Watch6. Over the 30 days preceding the second survey, between-group differences were significant by Welch's t-test (p = 0.0086), and discrimination reached AUC 0.859 with a 95% confidence interval of 0.664–0.992. In 90-day monitoring, generalized estimating equations (GEE) showed persistent divergence when groups were defined by the second survey, with significant differences across multiple weeks preceding the survey, consistent with the retrospective GAD-7 window. These findings indicate that ECG-pretrained cardiac representations can be translated to smartwatch PPG to support unobtrusive, continuous anxiety tracking in stroke. Results constitute feasibility-level evidence and motivate multi-site external validation, larger cohorts with clinician-rated assessments, and prospective studies toward clinical deployment.

Artificial intelligence (AI), particularly deep learning, continues to advance in medical image analysis and clinical prediction. In stroke care—where timely, accurate decisions are critical—AI is seen as a promising tool, with potential to detect complex imaging patterns and enhance clinical workflows. However, practical application in real-world settings remains limited, often due to structural issues in model design, evaluation, and insufficient integration of clinical context. This narrative review examines common pitfalls in developing and applying AI models in stroke care. High performance alone does not ensure clinical value; what matters is whether the predicted target (label) is clinically meaningful and well-defined. If the label is ambiguous or fails to reflect the underlying clinical condition, even highly accurate models may produce misleading or unhelpful outputs. We also discuss limitations of current explainability tools and emphasize that lack of interpretability hinders trust and adoption in high-stakes decisions. Rather than functioning as autonomous decision-makers, AI models are better positioned as coordinators or accelerators—supporting, not replacing, clinical judgment. For responsible integration into practice, developers must disclose key aspects of the model, including training data, label definitions, and performance conditions. Clinicians, in turn, should be prepared to interpret evaluation metrics in the context of real-world care. Ultimately, clinical AI should focus not merely on maximizing performance but on solving problems relevant to clinical practice, with transparency and explainability as essential prerequisites for adoption.

The presence of a diffusion-weighted imaging (DWI)–fluid-attenuated inversion recovery (FLAIR) mismatch holds potential value in identifying candidates for recanalization treatment. However, the visual assessment of DWI–FLAIR mismatch is subject to limitations due to variability among raters, which affects accuracy and consistency. To overcome these challenges, we aimed to develop and validate a deep learning-based classifier to categorize the mismatch. We screened consecutive acute ischemic stroke patients who underwent DWI and FLAIR imaging from a four stroke centers. Two centers were used for model development and internal testing (derivation cohort), while two independent centers served as external validation cohorts. We developed Convolutional Neural Network-based classifiers for two binary classifications: DWI–FLAIR match versus non-match (Label Set I) and match versus mismatch (Label Set II). A total of 2369 patients from the derivation set and 679 patients from two external validation sets (350 and 329 patients) were included in the analysis. For Label Set I, the internal test set AUC was 0.862 (95% CI 0.841–0.884, with external validation AUCs of 0.829 (0.785–0.873) and 0.835 (0.790–0.879). Label Set II showed higher performance with internal test AUC of 0.934 (0.911–0.957) and external validation AUCs of 0.883 (0.829–0.938) and 0.913 (0.876–0.951). A deep learning-based classifier for the DWI–FLAIR mismatch can be used to diminish subjectivity and support targeted decision-making in the treatment of acute stroke patients.

Chun MY, et al. J Clin Neurol. 2025 Jul;21(4):315-324. https://doi.org/10.3988/jcn.2024.0526

This study aimed to develop and validate an automated machine learning (ML) system that predicts 3-month functional outcomes in acute ischemic stroke (AIS) patients by combining clinical and neuroimaging features. Functional outcomes were categorized as unfavorable (modified Rankin Scale ≥ 3) or not. A clinical model employing optimal clinical features (Model_A), a convolutional neural network model incorporating imaging data (Model_B), and an integrated model combining both imaging and clinical features (Model_C) were developed and tested to predict unfavorable outcomes. The developed models were compared with each other and with traditional risk-scoring models. The dataset comprised 4147 patients from a multicenter stroke registry, with 1268 (30.6%) experiencing unfavorable outcomes. Age, initial NIHSS, and early neurologic deterioration were identified as the most important clinical features. The ML model prediction achieved an area under the curves of 0.757 (95% CI 0.726–0.789) for Model_A, 0.725 (95% CI 0.693–0.755) for Model_B, and 0.786 (95% CI 0.757–0.814) for Model_C in the test set. The integrated models outperformed traditional risk-scoring models by 0.21 (95% CI 0.16–0.25) for HIAT and 0.15 (95% CI 0.11–0.19) for THRIVE. In conclusion, the integrated ML system enhanced stroke outcome prediction by combining imaging data and clinical features, outperforming traditional risk-scoring models.

Background and aims: Predicting the prognosis of acute ischemic stroke (AIS) is crucial in a clinical setting for establishing suitable treatment plans. This study aimed to develop and validate a machine learning (ML) model that predicts the functional outcome of AIS patients and provides interpretable insights.Methods: We included AIS patients from a multicenter stroke registry in this prognostic study. ML-based methods were utilized to predict 3-month functional outcomes, which were categorized as either favorable [modified Rankin Scale (mRS) ≤ 2] or unfavorable (mRS ≥ 3). The SHapley Additive exPlanations (SHAP) method was employed to identify significant features and interpret their contributions to the predictions of the model.Results: The dataset comprised a derivation set of 3,687 patients and two external validation sets totaling 250 and 110 patients each. Among them, the number of unfavorable outcomes was 1,123 (30.4%) in the derivation set, and 93 (37.2%) and 32 (29.1%) in external sets A and B, respectively. Among the ML models used, the eXtreme Gradient Boosting model demonstrated the best performance. It achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.790 (95% CI: 0.775–0.806) on the internal test set and 0.791 (95% CI: 0.733–0.848) and 0.873 (95% CI: 0.798–0.948) on the two external test sets, respectively. The key features for predicting functional outcomes were the initial NIHSS, early neurologic deterioration (END), age, and white blood cell count. The END displayed noticeable interactions with several other features.Conclusion: ML algorithms demonstrated proficient prediction for the 3-month functional outcome in AIS patients. With the aid of the SHAP method, we can attain an in-depth understanding of how critical features contribute to model predictions and how changes in these features influence such predictions.