jakobkolb / ml-covid-nowcasting Goto Github PK

Sex: 1 female, 2 male

What about 1.5? ;)

The current feature construction from rhr data does not contain any information about infection according to feature importance. Use time lagging and other kinds of feature construction (e.g. catch22 feature library) to extract more info.

normalize rhr/step metrics by std

To make the claim that this model works with test capacity maxed out, we have to plot results from data without test results separately and compare them to epidemic dynamics.

Currently, the model seems to make reasonable predictions for omicron but has a lot of false positives for delta. This could potentially be explained by 1) the imbalance between omicron (many) and delta (few) cases in our data and 1) the relatively mild symptoms of omicron. So the model learns that most of the cases have symptoms like omicron and then has a lot of false positives for delta, where the symptoms are more severe.

This could be mitigated by telling the model about the prevalence of different variants and then using 2nd order polynomial features, so the model can fit separate sets of parameters for different variants.