Our customer had an ambitious goal to contribute to flattening the curve in a world-leading state regarding the effective vaccination campaign, Israel. As part of the larger project, our task was to develop a solution predicting the spread and infection risk of COVID-19.We have faced a triple challenge — the velocity of the disease spread, unexpected changes in the environment, and precise predicting up to the city district. Having neural networks and deep learning techniques in our arsenal, we took on the challenge.

Here is our way of tackling tasks like this:

• The best option to deliver precise prediction and accurate generalization in an uncertain environment is the artificial recurrent neural network (RNN), namely long short-term memory (LSTM). It mainly comes in handy in the state when the situation changes dynamically. LSTM preserves "long-term" memory, beneficial for time-series data (i.e., data listed in time order) describing infected rates.
• Normalizing data both for epidemic's beginning and real-time prediction could become a problem, meaning the massive difference between forecasts drawn on the different stages of the epidemic's development. However, the model learned how to deal with the statistical error at every particular stage of the disease spreading, as we have developed special techniques to implement that.
• As our model was intended to predict infection rates for a particular town or its district, we added embedding layers to the model. Embedding layers are "trying" to compress data on cities in a most representative way for a machine learning (ML) model. This function "taught" the ML model to seek interaction within all the data, providing insights per any location.

We created a solution for predicting epidemic development in a particular country, specifying an accurate forecast per geographical location (e.g., town), scaling up to cities' districts. The model's prediction scope includes around 300 towns in Israel and its sections in the cities, considering the scales' differences.