ELLIS against Covid-19

Extreme domain adaptation: an early COVID-19 risk score

ellis Uri Shalit 06 May 2020 - 06 May 2020
no image

06 May 2020 • 13:55 - 14:05

Uri Shalit (Technion)

Clalit health services is the healthcare provider for 4.5 million people in Israel. In the middle of March, with the pandemic starting to spread in Israel, it urgently needed to build a COVID-19 risk score. The score was meant to be used to warn people at risk, offer special “at-home” health services, and prioritize testing and care. The main challenge was the dearth of data: no datasets or scores have been published at that time, and the only data available was univariate marginals from China. This data was in the form of probability of death conditioned on age groups and comorbidities, separately. We developed a score by bootstrapping an existing Israeli flu risk-model to the univariate marginals from China: this required creating an approximate multivariate model for China, and adjusting the Israeli flu-model to match the marginals from China. The risk model we created was deployed and widely used in Israel. We are currently analyzing detailed Israeli patient data in order to create much more accurate models.


Questions & Answers

Link to the recording of the live Questions & Discussion session for this talk. 

  • Q: Why a linear probability model rather than a generalised linear model? Or at least transforming probabilities to be in -\infty,\infty.

    • A: For linear model we only need second order probabilities from China, for other models we would have needed the full p(x_1,...,x_d) which was unreasonable.

  • Q: Do you make your data/models available?

    • A: Unfortuantely we cannot share our data - I will ask about the model. Feel free to follow up with me about this, urishalit@technion.ac.il

  • Q: Can the model be recalibrated online while more data arrives?

    • A: There are by now enough COVID-19 patients and deaths in Israel such that the current model which is in use is based only on Israeli data. It is updated and recalibrated as more data arrives.


Thumb ticker uri shalit