RESUMO
A virtual standardized patient (VSP) prototype was tested for natural language understanding (NLU) performance. The conversational VSP was evaluated in a controlled 61 subject study over four repetitions of a patient case. The prototype achieved more than 92% appropriate response rate from naïve users on their first attempt and results were stable by their fourth case repetition. This level of performance exceeds prior efforts and is at a level comparable of accuracy as seen in human conversational patient training, with caveats. This level of performance was possible due to the use of a unified medical taxonomy underpinning that allows virtual patient language training to be applied to all cases in our system as opposed to benefiting a single patient case.