Help reduce covid-related death and to help the management of covid patients in the hospital by prediction for the patient risk from cheap, early and easy accessible metrics.
What it does
Receives input in the form of clinical statistics or an x-ray image and predicts whether the patient's condition will worsen or get better.
Predicts the patient risks in two levels
- Death Also, it delivers the features of importance for each feature.
How we built it
Parts in Jupyter Notebook and parts in python. The data was cleaned and features were added to bolster the model's accuracy. The death/recovery models were built in sklearn and had their hyperparameters optimized by the TPOT genetic algorithm. The image classification model was build in keras using a CNN.
Challenges we ran into
Mostly we are not sure that the data covers exactly everything we would like to know. For CNN in particular, we faced time and resource constraints and unfortunately could not train a very large sample set (n=94)
Accomplishments that we're proud of
We created a predictive pipeline to determine the outcome of a person's current covid condition. Our death prediction model achieved 95% accuracy and 95% recall while our CNN achieved 100% validation accuracy and recall on our sample size! We are proud of the collaboration, which included people from all over the world, where people came together to solve an important topic.
What we learned
- How to choose the model target based on the data we have
- How to assess scientific studies for feature importance
- Importance of clinical expertise in the team
- Importance for the model to be relevant and practically possible
- How to choose modeling approach
- The approach is a pilot, to finalize any product more effort will be needed than to work beside your full-time job.
Most of all we learned that the age and lab tests taken in early emergency visit are more important for covid related death prediction than the vital measurements.
What's next for Data Science Queens
Creating models trained on more data in order to more effectively predict covid outcomes for patients.