Atlas

Inspiration

In recent events, Hurricane Milton and Hurricane Helen severely impacted the state of Florida and its emergency services. We thought about what might have happened had the public known which hospitals would have the most patients, wouldn't the government and hospitals have been able to better prepare? Wouldn't recovery after the storm be easier if we knew how many new patients might come in one or two weeks after the storm? To this end, we realized we could use the power of AI to predict a map of hospital loads, we can create an Atlas.

What it does

Atlas is a platform designed to assist in predicting hospital load during and after hurricanes. This information is then displayed on a dashboard along with an informational map displaying the hurricane and its current predicted path overlaid onto the hospitals in the region.

How we built it

We utilized WatsonX AI to help build our main CNN for our prediction analysis. We also leveraged the platform's generative AI capabilities to provide helpful information for our users. To build the frontend, we used React to handle state and Mui/joy component library for UI. On the backend, we utilized Python and the FastAPI library to create a server that handles our API calls and data parsing.

Hospital Use Prediction And Analysis

We used IBM's watsonx platform to help us develop a project we can be proud of. IBM's watsonx platform was integral in the development of our convolutional neural network (CNN) that predicts the hospital encounters for each week, from week 0, when a hurricane hits, to week 7. We used IBM’s watsonx platform to deploy MistralAI to perform the analysis of the results the CNN predicts. To train our model we used data from the Healthcare Cost and Utilization Project (HCUP) detailing hurricane effects on hospital use and our model takes in the county population a hospital resides in, the hospital proximity to the hurricane, and the category of the hurricane to deliver our predictions.

Challenges we ran into

The development of the map utilized geojson data which we have no previous experience with and so we had to adapt rapidly to handle this data along with the calculations to determine whether a hospital was in the direct path of a hurricane or otherwise for predictive model purposes. Despite the lack of experience in building CNN, our team managed to pull through on delivering an effective model through the use of AI to create the model which helped supplement the inexperience in packages like Pytorch.

Accomplishments that we're proud of

Storm Map

We developed a storm map that displays the path of the storm and the hospitals in the region.

AI Powered Analysis of Results

Using MistralAI through IBM's watsonx platform we were able to create analyses based off of the CNN's predictions to display to users.

Prediction Accuracy

In our tests of Hurricane Milton's impact on Pinellas county hospitals and care centers, we predicted consistently with >95% accuracy the total Pinellas county hospital encounters using our CNN. The total number of patients who were ordered to evacuate from the county and its hospitals was 6,600 and on average our predictions aligned with this number.

What we learned

Our team gained experience with IBM's watsonx platform and the value it can provide in augmenting a programming team. We learned about the utilization of Pytorch and CNNs along with geojson file utilization.

What's next for Atlas

The next step is expansion and scale. Our current model can always improve with more data and examples. We can further diversify our model by predicting exactly the injuries caused by a storm and we can expand to predictions on different disasters. Pushing Atlas to become an accurate model for not just Florida but all of the United States or even the globe can be considered an ultimate goal for the platform.

Built With

• ai
• cnn
• ibm-watson
• javascript
• jupyternotebook
• machine-learning
• python
• react