Caring for hospitalized COVID patients can be resource-intensive.  The Fuse Breakers seek to help identify patients who will require high flow oxygen in 96 hours of admission.  We collected over 300 hospitalized COVID patient data from July to August of 2020 in a local hospital.  We created a computer model using Xgboost, a machine learning program with decision tree feature, to forecast which hospitalized patient will need advance oxygen support (high flow, BiPAP, CPAP, Intubation) in 48 hours of admission. The accuracy is further boosted by five fold cross validation.  The model is available for review here. Our study shows that close to 45% of new COVID19 patients presented to our local hospital will require advanced oxygen support within 96 hours of admission.  The local hospital admin used this information to assist with their staffing assignments. 

We collaborated with programmers, medical students, and physicians to create this model.

In keeping up with the President Trump’s space themed Project Warp Speed, we code name this program as the “Apollo Project” and name our models after Mars’ rovers.

The system uses an ensemble methodology, with 5 separate models making independent calculations and then vote on the prediction. 

The model reached a respectable accuracy profile.

We further tested the model with additional patient data not seen by the computer during training phase. 

The model yielded good results

Here is a video of our computers working hard to create the 5 models used in this system. 

Surprisingly while age is a factor that correlates to future oxygen need, it is not the strongest factor.  Inflammatory markers like CRP, ESR, Ferritin, LDH, and WBC have higher correlation.  The initial oxygen requirement is also a good predictor of future oxygen need.  

There is an articled published by the Annals of Internal Medicine on the September 22, 2020  that share share similar data for hospitalized COVID19 patients.   Please note these models are for statistical analysis only.  It is not intended to affect treatment decisions.  Do not base any treatment from this model.