The Autism and Developmental Disabilities Monitoring (ADDM) Network conducts population-based surveillance of autism spectrum disorder (ASD) among 8-year old children in multiple US sites. To classify ASD, trained clinicians review developmental evaluations collected from multiple health and education sources to determine whether the child meets the ASD surveillance case criteria. The number of evaluations collected has dramatically increased since the year 2000, challenging the resources and timeliness of the surveillance system. We developed and evaluated a machine learning approach to classify case status in ADDM using words and phrases contained in children's developmental evaluations. We trained a random forest classifier using data from the 2008 Georgia ADDM site which included 1,162 children with 5,396 evaluations (601 children met ADDM ASD criteria using standard ADDM methods). The classifier used the words and phrases from the evaluations to predict ASD case status. We evaluated its performance on the 2010 Georgia ADDM surveillance data (1,450 children with 9,811 evaluations; 754 children met ADDM ASD criteria). We also estimated ASD prevalence using predictions from the classification algorithm. Overall, the machine learning approach predicted ASD case statuses that were 86.5% concordant with the clinician-determined case statuses (84.0% sensitivity, 89.4% predictive value positive). The area under the resulting receiver-operating characteristic curve was 0.932. Algorithm-derived ASD "prevalence" was 1.46% compared to the published (clinician-determined) estimate of 1.55%. Using only the text contained in developmental evaluations, a machine learning algorithm was able to discriminate between children that do and do not meet ASD surveillance criteria at one surveillance site.