Model

Build and Run Docker Contrainer

The Docker container can be build with the following command:

$ docker image build -t triage_predictor .

Finally, you can run the docker container with:

$ docker run -p 5000:5000 triage_predictor

Prediction Call to Flask

You can get a prediction from the Triage model, by sending it a JSON in the following format:

{ 
  "patientID" : ... , #string
  "heart_beat" : ..., #int
  "oxygenation" : ..., #float
  "temperature" : ..., #float
  "breathing_rate" : ..., #int
  "preconditions" : ..., #string (None, Hypertension, Asthma, Arthritis, Cancer)
  "fitness" : ..., #int [1,10]
  "smoker" : ... #bool
}

$ curl --header "Content-Type: application/json" \
  --request POST \
  --data '{"patientID" : "111111", "heart_beat" : 50, "oxygenation" : 0.99, "temperature" : 36.5, "breathing_rate" : 18, "preconditions" : "None", "fitness" : 9, "smoker" : false }' \
  http://0.0.0.0:5000/predict

Data

Input data:

• Patient - clinician Data:

  • name (str)
  • surname (str)
  • Location (categorical)
  • Patient Unique Id (uuid)
  • Doctor Unique Id (uuid)

• When the patient was added to the system:

  • Timestamp (timedt)

• Medical parameters (not a timepoint, it is a history of values):

  • timepoint
  • oxygenation level (percentage)
  • heart beat (real)
  • patient age (int)
  • pregressed conditions (categorical)
  • temperature (float)
  • number of days the symptoms have started (int)
  • breathing rate (float)

Output data:

• Triage level, for now we set 5 levels of severity. 1 least severe, 5 most severe. More input from clinicians will be required.
• Confidence (%)

Model and plan

Assumptions:

• not all the features counts the same (breathing rate or oxygenation and age should be the most relevant), so we have to weight them before applying the model.
• for each feature we have a collection of time-points. The last time-points are the most important ones in making a prediction.

Plan:

• Train a random forest (1) to weight the features
• Non-linear mixed model or regression to get the trend from the historical data (extra weight on latest timepoints).
• Combine latest timepoint and trend in a single input data.
• Train another random forest (2) to classify the combined data above and to get a prediction.

Random forest 1 and 2 should be trained on different datasets.