Pediatric acute myeloid leukemia (AML) is an aggressive blood cancer with a poor prognosis and high relapse rate. Current challenges in the identification of immunotherapy targets arise from patient-specific blast immunophenotypes and their change during disease progression. To overcome this, we present a new computational research tool to rapidly identify malignant cells. We generated single-cell flow cytometry profiles of 21 pediatric AML patients with matched samples at diagnosis, remission, and relapse. We coupled a classifier to an autoencoder for anomaly detection and classified malignant blasts with 90% accuracy. Moreover, our method assigns a developmental stage to blasts at the single-cell level, improving current classification approaches based on differentiation of the dominant phenotype. We observed major immunophenotype and developmental stage alterations between diagnosis and relapse. Patients with KMT2A rearrangement had more profound changes in their blast immunophenotypes at relapse compared to patients with other molecular features. Our method provides new insights into the immunophenotypic composition of AML blasts in an unbiased fashion and can help to define immunotherapy targets that might improve personalized AML treatment.