Introduction: Stress perfusion cardiac magnetic resonance (CMR) is a guidelines-backed non-invasive test for the assessment of coronary artery disease (CAD), and its prognostic value is well validated. However, image interpretation requires a high level of expertise. Furthermore, the direct relationship between image pixels and outcome is not well understood. Hypothesis: Qualitative stress perfusion map based on American Heart Association (AHA) segmentation and direct linkage with outcome can be achieved using artificial intelligence. Methods: Retrospective CMR data collection was performed in one center between 2011 and 2021. Cases with death events were identified. Mean time to event was considered as a follow-up period, and cases without events and with shorter duration from CMR to collection date were excluded. Stress perfusion images were selected from 3 left ventricular (LV) levels: basal, mid and apical. Clinical CMR reports with binary AHA perfusion values were available. Six convolutional neural networks (CNNs) were trained to predict AHA segments values for each level and merged into four CNNs for the apical level. Perfusion, scar image pixels and clinical data were transformed into features, and a hybrid CNN (HNN) was trained to predict mortality. Training, validation, and test split was 70%, 15%, and 15%, respectively. Evaluation metrics included area under curve (AUC). Results: Total number of cases included for AHA plot classification were 2139 (6417 perfusion images). The average AUC for AHA perfusion classifiers was 61%, an example with correct predictions is presented. Follow-up period was calculated at 999 days. Total number of cases included for HNN development was 1294. HNN achieved high prediction value with AUC of 81%. Conclusions: Qualitative AHA stress perfusion map was feasible using artificial intelligence. Outcome prediction with HNN was a powerful new approach for risk stratifying patients with known or suspected CAD.