Although clinical decision support systems can reduce costs and improve care, the challenges associated with manually maintaining content has led to low utilization. Here we pilot an alternative, more automatic approach to decision support content generation. We use local order entry data and Bayesian networks to automatically find multivariate associations and suggest treatments. We evaluated this on 5044 hospitalizations of pregnant women, choosing 70 frequent order and treatment variables comprising 20 treatable conditions. The method produced treatment suggestion lists for 15 of these conditions. The lists captured accurate and non-trivial clinical knowledge, and all contained the key treatment for the condition, often as the first suggestion (71% overall, 90% non-labor-related). Additionally, when run on a test set of patient data, it very accurately predicted treatments (average AUC .873) and predicted pregnancy-specific treatments with even higher accuracy (AUC above .9). This method is a starting point for harnessing the wisdom-of-the-crowd for decision support.