Decision Support in Health Care via Root Evidence Sampling

Bayesian networks play a key role in decision support within health care. Physicians rely on Bayesian networks to give medical treatment, generate what-if scenarios, and other decision-support tasks. Stochastic sampling from a Bayesian network with some nodes instantiated as evidence is a powerful tool with Bayesian networks. With decision support systems, generating random samples from a Bayesian network is key to simulating possible scenarios and consequences. Some techniques for stochastic sampling, such as logic rejection sampling or importance sampling, can be very slow when given unlikely evidence. We propose root evidence sampling (RES), an algorithm that carefully reorganizes some or all of the evidence nodes to be root nodes, computes new conditional probability tables, and then uses simple forward sampling or a hybrid approach to generate samples. We show that RES performs favorably compared to other sampling techniques without sacrificing accuracy, particularly when the evidence is unlikely. We also show that the new network generated by RES has the same inferential capability as the original network, which has implications for structure learning