Seminar: Caroline Jagtenberg (VU Amsterdam)

Zoom link: 
https://cwi-nl.zoom.us/j/84909645595?pwd=b1M4QnNKVzNMdmNSVFNaZUJmR1kvUT09 
(Meeting ID: 849 0964 5595, Passcode: 772448)

Speaker: Caroline Jagtenberg (VU Amsterdam)

Title: Analytics for community first response

Abstract:
In Community First Responder systems, traditional ambulance response is augmented by a network of trained volunteers who are dispatched via an app. A central application of such systems is cardiac arrest, where shortening the time to CPR is crucial. A number of such community first responder (CFR) systems are active worldwide, for example HartslagNu in the Netherlands and GoodSAM in the UK, Australia and New Zealand. We have received detailed data from the latter, which we combine with mathematical modeling to optimize the following:

1) Recruitment. For a target performance level, how many volunteers are needed and from which locations should they be recruited? We model the presence of volunteers throughout a region as a Poisson point process, which permits the computation of the response-time distribution of the first-arriving volunteer. Combining this with known survival-rate functions, we deduce survival probabilities in the cardiac arrest setting. We then use convex optimization to compute an optimal location distribution of volunteers across the region.

2) Phased alerts. GoodSAM notifies volunteers in batches with built-in time delays. The policy that defines the batch sizes and delays affects the time to CPR as well as the number of redundant volunteer arrivals. We start by estimating these KPIs for various policies through Monte Carlo Simulation. We continue by using machine learning to predict the best policy to use, given where the volunteers are observed in relation to the patient. We do this by formulating the problem as a multiclass classification problem, for which we train a tree on the results from the simulations above. We compare the performance of the tree against a policy designed by dynamic programming. Finally, we look into optimal decision trees which go beyond the heuristic nature of machine learning algorithms.

Slides:
Slides