Autism Spectrum Disorder is a highly heterogeneous condition currently diagnosed using behavioral symptoms. A better understanding of the phenotypic subtypes of autism is a necessary component of the larger goal of mapping autism genotype to phenotype. However, as with most clinical records describing human disease, the phenotypic data available for autism contains varying levels of noise and incompleteness that complicate analysis. Here we analyze behavioral data from 16,291 subjects using 250 items from three gold standard diagnostic instruments. We apply a low-rank model to impute missing entries and entire missing instruments with high fidelity, showing that we can complete clinical records for all subjects. Finally, we analyze the low-rank representation of our subjects to identify plausible subtypes of autism, setting the stage for genome-to-phenome prediction experiments. These procedures can be adapted and used with other similarly structured clinical records to enable a more complete mapping between genome and phenome.