Google Introduces PPE, a System That Collects Its Own Data for Geographic Forecasting
Google Introduces PPE, a System That Collects Its Own Data for Geographic Forecasting
On 27 August, a Google research team introduced the Planetary Prediction Engine, a system that takes a task described in text and labeled source data, then produces a geographic forecast.
To decide which districts to monitor the following week, an epidemiologist needs more than an algorithm. They must find data on the local environment and population, align the data to consistent geographic boundaries and dates, and exclude information that reveals the outcome in advance. The authors of the technical preprint estimate that preparing an operational epidemic forecast requires more than 700 actions.
PPE first identifies the geographic area and time period covered by the task and determines how the tables should be joined. The system then searches open geographic databases, government portals, and scientific repositories for suitable information. It aligns the data to consistent district boundaries and assembles a training dataset. PPE uses this dataset to build a forecast, which it evaluates on separate districts and weeks.
During the same outbreak, the PARTNERS trial is enrolling patients after Bundibugyo virus disease has been confirmed. PPE identifies districts for monitoring before their first recorded cases appear.
During the outbreak caused by Bundibugyo ebolavirus in the Democratic Republic of the Congo, PPE ranked districts with no recorded cases once a week. Across five weekly forecasts, the system placed 15 of the 18 districts that recorded their first cases the following week among the ten highest-risk districts, which corresponds to 83.3%. The preprint authors compare this result with approximately 73% for a previously published model that also ranked ten districts by risk.
PPE keeps model training separate from forecast evaluation. The Feature Gate filter removes features that reveal the outcome in advance, including components of the variable being predicted, data from the same questionnaire, consequences of the event, and information from the future. When values are missing, the system calculates the values used to fill them from the training data alone and then applies them to the evaluation data. Districts and weeks reserved for evaluation do not contribute to model preparation.
The system then tests several model types, evaluates them across different geographic areas and time periods, and produces a report listing districts in order of risk.