Recorded on February 23, 2018.
In this webinar, we wish to explore how transmitting additional information about the nature of an accident can help dispatchers request rescue services that are adequate for the anticipated severity of injuries. For this purpose, we examine accidents recorded in NHTSA's National Automotive Sampling System (NASS), for which a wide range of accident attributes are available, including the injury severity of passengers of vehicles involved in the accidents.
As it turns out, predicting injury severity is a fairly straightforward task, which can be achieved by a wide range of modeling techniques with good overall classification performance. Much more difficult is determining the degree of confidence one should have in a forecast for a particular accident instance. Needless to say, the incorrect provisioning of rescue services is a matter of life and death.
Modeling this problem domain with Bayesian networks offers many advantages. Most importantly, Bayesian networks are inherently probabilistic, which means that any forecast is in the form of a probability distribution. If, for instance, an individual injury forecast provided an expected value of "minor injury", we would also see a value for the risk of "severe injury." Thus, services could be provisioned on the basis of the risk of a severe injury as opposed to the average of the injury forecast. This allows for a very formal approach of dealing with resource constraints related to overprovisioning and the dangers of underprovisioning.
In this context, another benefit of inference with Bayesian networks becomes apparent: Bayesian networks permit incremental inference. Despite the benefits of automatic collision notification, one would still expect missing data, redundant information, conflicting and uncertain evidence emerge from emergency situations. Bayesian networks can handle any newly-emerging information and compute an updated forecast as often as needed.