Relapse in acute myeloid leukemia (AML) may result from variable genetic origins or convergence on common biological processes. Exploiting the specificity and sensitivity of regulatory DNA, we analyze patient samples of multiple clinical outcomes covering various AML molecular subtypes. We uncover regulatory variation among patients translating into a transcriptional signature that predicts relapse risk. In addition, we find clusters of coexpressed genes within this signature selectively link to relapse risk in distinct patient subgroups defined by molecular subtype or AML maturation. Analyzing these gene clusters and the AML subtypes separately enhances their prognostic value substantially and provides insight in the mechanisms underlying relapse risk across the distinct patient subgroups. We propose that prognostic gene expression signatures in AML are valid only within patient subgroups and do not transcend these subgroups.