An alert system for a third wave of Covid-19 cases in South Africa has been developed using Artificial Intelligence (AI) to uncover complex interdependencies and relationships between mobility data, stringency of non-pharmaceutical interventions (NPIs) and actual case data. This alert system recognises an additional wave when the actual value of daily cases deviates from the predicted value for each province. The risk metric used for the alert system is the relative difference. The Risk Index (the relative difference), is calculated using the following equation:
Artificial Intelligence Algorithm: LSTM RNN
The prediction model is based on a Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM). Though recurrent neural networks are often used in time series forecasting, they have a problem where the weights of past data are lost due to the vanishing gradient problem. Adding Long Short Term Memory to the RNN architecture is the solution to this problem.
RNNs, like other neural networks, have an input layer, a hidden layer and an output layer, but are characterised by the fact that the hidden layer changes under the influence of the previous time's hidden layer and the new input layer. The figure below shows the structure of a RNN where x is the input layer, h is the hidden layer and y is the output layer:
LSTM RNNs have the same basic structure as simple RNNs. However, while RNNs use a simple function to generate the hidden layer, LSTM uses a more complex function to create the hidden layer for long-term memory.
The predictions were based on the following features. (Please click on the links if you want to know more information.)
Actual positive cases are updated every day, while Google Mobility and Facebook Mobility Indices are updated weekly with a delay of a week.
Based on the relative differences between the predicted and measured values at steady-states between the first and the second waves, the threshold for the possible onset of the third waves were determined as follows:
The predictions are updated every Monday, and the Risk Index is updated every day.
Users should concentrate on the red line on the graphs below. The red line should be below the orange line (the risk threshold). When the red line is consistently above the orange line, the probability of a wave is high. It must also be noted that a low index can be an indication that a province has past a wave.