MLEDGE — Epidemic-risk forecasting from telco mobility

A complete picture of epidemic risk in real time would require something governments don’t have: mobility data from every telco serving a country, processed together. The textbook obstacle is privacy — both individual privacy and the commercial sensitivity of one operator’s customer base to another. MLEDGE’s digital-economy use case set out to demonstrate that the picture can be assembled without ever centralising the data, using federated analytics over real telco records.

Working with Orange España (37 million active lines), Acuratio led the design and implementation of the federated-learning components and the ML models behind the risk computation. The team used Call Detail Records and network-probe events to reconstruct anonymised user trajectories across Madrid and Barcelona for the first two weeks of March in 2020, 2021, and 2024 — capturing the COVID peak, the year after, and a post-pandemic baseline. All locations are encoded as Geohashes, so that nearby points share prefixes — which makes federated aggregation tractable.

From this data the system infers each user’s home (night-time CGI) and workplace (10:00–14:00 weekday CGI), then layers two clinically relevant detections on top: hospitalisations, identified by users co-located with a hospital geohash for at least four consecutive nights, and self-confinement, identified by users staying more than 10 hours at home for at least four consecutive days. On a representative day in March 2024 in Madrid, the pipeline detected 7,593 users spending more than 10 hours in a hospital and 1,315 hospitalised for more than four consecutive days, drawn from a base of 1.1 million users with full home/work information. Cross-referencing with hospital-bed capacity turned the raw counts into an occupancy map that highlights centres approaching saturation — the kind of view an emergency-management team can act on.

To prove federation works in practice, the same analyses were re-run after splitting the data two different ways: by user id (simulating multiple operators) and by territory (simulating regional providers). In both cases, the federated maps reproduce the centralised maps faithfully — including a useful failure mode the team caught and reported: a hotel cluster initially detected as a hospital, a reminder that anomalies in spatial models need urban-context validation.

The work was named a finalist at La Chambre’s annual gala in the “Innovation – Large Enterprise” category. MLEDGE is funded by the Spanish Ministry of Economic Affairs and Digital Transformation through EU NextGeneration-EU; the work was reviewed by IMDEA Networks.