My idea is to adapt an existing model for predictive maintenance. the following is a starting point for discussion, ideas, collaboration and possibly project work.
Please share your comments - email@example.com
Goal Classify patients as healthy vs. probability of being infected
Value Help healthcare planning, ICU planning
- Patient data, health records, symptoms, mortality, relevant
- pathologies (e.g. blood pressure, diabetes)
- Patient whereabouts
Classification Healthy vs. probability of requiring ICU within x hours
Leverage common features:
- Time-delay features: time series of observations leading to a critical case or death
- symptoms, vitals (e.g. body temperature, etc) similar to ‘telemetry data’
- number of days a person has been in danger areas
supervised ML model.
Labeling strategy: label as failed all cases within a time window in advance of failure. In the case of coronavirus the time window could be chosen as ‘sufficient lead time’ for the ICU
Algorithms random forests or other ML algorithms (adaboost?) maximize recall because the real goal is to reduce false negatives