ABSTRACT NUMBER: ESOC2026A1848 EXPLAINABLE DEEP LEARNING MODELS REVEAL WHITE MATTER HYPERINTENSITY VOLUME AS STRONG FUNCTIONAL OUTCOME PREDICTOR AFTER LARGE VESSEL OCCLUSION STROKE

Abstract

Abstract Background and aims Acute ischemic stroke (AIS) is a major global health burden that can potentially benefit from accurate prognostic strategies. Deep learning approaches, particularly convolutional neural networks (CNN), have demonstrated good predictive performance but are limited by lack of interpretability and uncertainty at the individual patient level. We aim to improve stroke outcome prediction by deriving relevant features from acute-phase stroke imaging. Methods We analyzed 1’024 patients with AIS from two separate stroke image repositories, together with routinely acquired imaging and clinical data. A CNN was trained on diffusion-weighted MRI to predict 3-month functional outcome dichotomized by the modified Rankin Scale (0-2 vs. 3-6). Model performance was evaluated using 10-fold cross-validation. Gradient-weighted class activation mapping (Grad-CAM) was applied to identify imaging features relevant for prediction, which were subsequently assessed for their association with outcome. Results The CNN achieved a moderate prediction performance with a mean AUC of 0.72 (standard deviation: ± 0.1), in line with previous imaging-based models. Grad-CAM analysis identified three three salient image features: brain volume relative to total intracranial volume, white matter hyperintensity (WMH) volume and ischemic lesion volume. A logistic regression model based only on WMH volume achieved an AUC of 0.71. Conclusions WMH volume, a surrogate of biological brain age or “brain frailty”, was strongly associated with functional outcome and demonstrated a predictive performance on par with the CNN. Future models may benefit from focusing on outcomes closer to acute phase interventions, such as early neurological recovery or discharge mRS. Conflict of interest Hakim Baazaoui received funding from the Koetser Foundation and the «Young Talents in Clinical Research» program of the SAMS and of the G. & J. Bangerter-Rhyner Foundation. Pascal Bühler: nothing to disclose. Julian Deseö: nothing to disclose. Martin Hänsel: nothing to disclose. Lisa Herzog: nothing to disclose. Beate Sick: nothing to disclose. Susanne Wegener received research funds by the Swiss National Science Foundation, (3100030_200703), the UZH Clinical research priority program (CRPP) stroke, the Zurich Neuroscience Center (ZNZ), the Baugarten foundation, the Hartmann Müller Foundation, the Koetser Foundation, the Philas Foundation, the Swiss Heart Foundation; and speaker honoraria from Amgen, Springer, Advisis AG, Teva Pharma, Boehringer Ingelheim, Lundbeck, Astra Zeneca, FoMF, and a consultancy fee from Bayer and Novartis via institution for research.

Ähnliche Arbeiten