Problem: current ambulance routing don’t optimize significantly on the contextual cases for stroke patients
Stroke hospitals: PSC is smaller than a CSC.
Previous work
Routing methods—
- route all patient to nearest PSC, which is worse than
- route high risk patient to CSC, which is worse than
- always route to CSC
This is counter-intuitive. How do we solve, given a stroke condition, available PSC/CSC locations, traffic, etc., for where and how to route a patient?
Ambulance MDP formulation
- \(S\): (location, symptom onset, known stroke type, stroke type)
- \(A\):
- route to clinic, route to [specific] PSC, route to [specific] CSC
- will never be downrouted (for instance, if you are at a PSC you will always either stay or go to CSC)
- \(T(s’|s,a)\):
- location changes
- distance
- \(R(s,a)\):
- “probability of patient outcome” \(P(success|time)\) (Holodinsky, et. al. 2018)
- if stroke type is unknown, its a weighted average
Solving
Forward Search, depth of 2: patient will either get transported or bounced and transported.
Results
- status quo: people near Stanford hospital/ChanZuck are better
- MDP: smoother gradient