Description as a Tweet:

Clairvoyant uses machine learning to predict where 911 calls for ambulances will come from. The data that is generated can then be used to route a nearby ambulance to the location soon before the call goes out.

Inspiration:

During COVID-19, the volume for emergent ambulance calls has skyrocketed, leaving many systems stressed to their breaking point. Preemptively sending an ambulance to a location before a 911 call goes out saves time for everybody involved. The patient is able to receive care faster, and the 911 system is given more flexibility since more ambulances will be available.

What it does:

Our project predicts where the next 911 call will come from, which allows ambulances to be preemptively dispatched. Our project predicts 911 calls in NYC.

How we built it:

We coded our project in Python. For the machine learning model, our group used Facebook Prophet. We decided on this model because Prophet is able to handle seasonality very well. For example, holidays often see an increase in the volume of 911 calls. For data processing and cleaning we used pandas and scikit-learn.

Technologies we used:

Python
AI/Machine Learning

Challenges we ran into:

The biggest challenge was how to represent features in our model. For example, ZIP codes are a high cardinality feature, and had to be encoded in order to be properly handled. However, encoding came with its own set of problems. Target encoding was identified as the best way to represent some high cardinality values, such as ZIP codes. However, encoding them like that makes the feature for the ZIP codes lose their spatial relationships that they have, which complicated things.

Accomplishments we're proud of:

Being able to create something that we knew very little about.

What we've learned:

We all learned a great deal about machine learning in general. We are all very new to the field, so the whole event was just trying different things and figuring out what works. Learning how a lot of different types of ML models work was also very interesting.

What's next:

Trying to improve the model further by combining it with different ML models that can take different types of encoding as input. Prophet is unfortunately limited in what can be used as input, and combining with a different model could help solve this. Creating a model that works in rural areas where resources are even more scarce would also make this project more impactful.

Built with:

Python, Facebook Prophet, pandas, and scikit-learn. The data came from: https://data.cityofnewyork.us/Public-Safety/EMS-Incident-Dispatch-Data/76xm-jjuj

Clairvoyant