Background: The prognosis of stroke patients with proximal vessel occlusion has significantly improved with the introduction of endovascular thrombectomy (EVT). In the extended time window beyond six hours, patient selection is primarily guided by perfusion-based CT imaging using thresholds. However, this approach may not fully capture the rich information contained within high-dimensional CT data. In contrast, advanced machine learning techniques have demonstrated their capacity to model complex relationships between multiparametric data, producing more precise predictions of the final infarct extent. Objective: Our study aims to develop and validate a deep learning approach that improves the prediction of final infarct based on the thrombectomy outcome. Methods: We conducted a multicenter retrospective study involving 405 stroke patients with acute proximal vessel occlusion in the anterior circulation who underwent EVT. To gain insights from this dataset, we developed a convolutional neural network (CNN) featuring a 3D multi-path hybrid-fusion architecture and attention units. The model leverages acute multimodal CT imaging as well as clinical and demographic characteristics. After being trained with knowledge of the thrombectomy outcome and final infarct localization as post-hoc assessed markers, our model can simulate the individual tissue outcomes of both successful and unsuccessful thrombectomy for new patients. The predictive accuracy of the CNN was benchmarked against a thresholding-based method as used in clinical practice and a generalized linear model (GLM). Results: By integrating multiparametric imaging data and patient-specific factors, our CNN enables individual voxel-based infarct prediction depending on thrombectomy outcome. When tested on an internal (n = 50) and an external dataset (n = 51), it outperformed both methodological baselines significantly in terms of spatial accuracy with Dice scores of 0.47/0.52 vs. 0.26/0.36 and 0.34/0.35 for the thresholding-method and the GLM, respectively Conclusion: Our deep learning model allows for a robust prediction of the final infarct extent as a function of thrombectomy success. This technique could potentially serve as a new biomarker for assessing the individual benefit of EVT in acute stroke care, thus supporting decision-making in the extended or unknown time window.