A visual analytics tool is helping the Gauteng government visualise prediction models provided by the University of the Witwatersrand (Wits) to see predicted active Covid-19 cases, as well as predicted hospital bed availability over time.

This can help determine where to focus resources ahead of when they are needed to ensure hospitals are able to cope with the expected number of cases, IBM Research Africa software engineer Richard Young says.

The platform was developed by IBM Research to help the Gauteng government track the spread of the virus into various areas of the country over time. It serves as a visual analytics tool where, instead of displaying statistics and charts, it allows users to drill down into areas of interest, analyse specific time periods and cross filter on a number of dimensions such as age group and gender.

The platform was unveiled by IBM Senior VP and Wits alumni Bridget van Kralingen, during the IBM Think conference held on June 24, with IBM Research Africa scientist Ashley Gritzman providing insight into its functionality.

The dashboard was designed to address three key questions for policy makers including where the current hotspots are, what the demographics and risk factors of the hotspots are and what the forecast for the spread of the virus based on the five levels of intervention is, says Young.

"For example, by using the dashboard, it is clear to see hotspots in March in more affluent areas as business travellers returned from overseas. However, these areas were eventually overtaken in April as the virus spread through communities to other districts. Such insights can help officials decide where health workers are needed and the type of intervention strategy [that is needed]."

A key component of the platform is being able to cross-reference the location of known Covid-19 cases with vulnerable areas using demographics and risk factors provided by the Gauteng City-Region Observatory (GCRO). Using this, the Gauteng government can monitor and focus its response to areas that are more susceptible to becoming hotspots owing to factors such as crowded residential areas or areas that lack access to appropriate water and sanitation, explains Young.

Wits University developed prediction models for each of the five alert levels for nonpharmaceutical intervention control strategies. Data analytics, such as these, can give officials the appropriate instruments to estimate how many beds and ventilators will be needed at the peak of the wave based on level of alert.

"Using the platform, officials can see a hotspot emerging in a densely packed ward. Understanding this will make it clear that enforcing physical distancing will be less effective and will also be useful in preparing more beds ahead of time."

Young adds that developers tried to make it intuitive for the system to answer complex questions such as: "Which areas have seen the highest number of new cases in the past week in the age group 60 to 70 years in high risk areas".

The queries are done by selecting options on each of the charts and the results are displayed in real-time on colour-coded heat maps.

Using the prediction models enables the platform to answer 'What-if' queries and questions like: "How will the spread of Covid-19 look if we change lockdown levels tomorrow" and allows the user to jump to specific dates to see the expected cases in each area. The platform can also specify a chain of interventions such as a series of lockdown levels and dates to see how it would affect the numbers, says Young.

"It can be quite complicated to fully understand all the features of a machine-learning and cloud-technology platform like this, and not all features will be relevant to all users. We have tried to provide sensible defaults for the features that many users will not need to use, while still allowing those features to be adjusted by users who need more advanced functionality," he notes.

DATA INSIGHTS
The Covid-19 pandemic has created vast amounts of geospatial and temporal data. Machine-learning can help by identifying patterns and insights in data that are not obvious, as well as provide predictions based on previous data.

The rapid increase in computing power in the cloud seen in recent years means users are able to run these complex algorithms in real-time and on bigger data sets. Policy makers can get real time predictions on the impact that various interventions would have, Young highlighted.

Additionally, the platform has been built to combine datasets from various sources and display them in a unified way. The dashboard uses anonymised private and public Covid-19 data from the Gauteng government and the University of Pretoria. The risk-factor data is provided by the GCRO.

"These are data sources that the Gauteng Government was already using before we unified them into our platform. We ensure that all data is sufficiently anonymised before it enters our platform by aggregating it across the dimensions that are useful for policymaking. We only use data from verified sources and help the data providers to check the accuracy of their data by detecting outliers or anomalies."

"During the week of June 29 to July 3, IBM also announced it had donated its AI Fairness 360 toolkit to the Linux Foundation. This toolkit helps developers examine, report, and mitigate discrimination and bias in machine learning models throughout the AI application," adds Young.