The aim of this study was to use functional profiling and AI to identify drug resistance and predict risk in patients with pediatric acute myeloid leukemia (pedAML). 45 pedAML patient samples of blood/bone marrow mononuclear cells from the biobank of St. Anna Children’s Hospital in Vienna, taken at diagnosis, were analyzed, alongside, an additional 5 samples that were used as benchmarkers. In the retrospective study, the cells were incubated for 24 hours, then stained with blast and T-cell specific markers and then imaged. The study used high-throughput imaging with custom analysis of cell markers to classify cells into blasts, T cells, and other cells and assess their viability. The model was able to classify cell type with an overall accuracy of 87% and cell viability with an accuracy of 92%. The model was further tested by performing image-based drug sensitivity profiling with a Pearson correlation score of 0.83. The sample cohort was chosen to cover the majority of molecular risk groups and this was used in tandem with drug profiling, multiomics and the relative blast fraction (RBF) to initially show overall no association with classical clinical parameters such as cellular identity as indicated by FAB class, cytogenetic subgroups, or initial response. Instead, the study found correlations between clusters of compounds demonstrating that the effectiveness of one agent can indicate whether other agents will be comparable or worse in similar patient populations. Further analysis through observations, a support vector classifier, and propensity scores showed that specific genes were associated with higher risk cell differentiation states (both monocyte-like and hematopoietic stem cell-like) which further influenced additional drug responses. Ultimately, machine learning models using RBF-AUC scores were able to predict measurable residual disease and early relapse with 85.2% and 79.2% respectively. Overall, the study demonstrates that ex vivo function profiling, including the use of AI, can be incorporated into precision medicine-based approaches to predict risk and identify targeted treatments early and throughout the disease course of pedAML. While the deep learning model shows promise, the small patient cohort, rarity of pedAML, and nature of retrospective studies do not allow for demonstration of the full capabilities of the image-based profiling model. The study also lacks a validation cohort, making it difficult to implement its findings. Overall, more work to improve the model alongside prospective studies that include larger patient cohorts is needed.