Greece deployed Eva, a reinforcement-learning system directing COVID-19 tests at high-risk incoming travellers across 40+ border points, detecting 2× more asymptomatic cases than random testing while processing over 40,000 households daily.
2x
Asymptomatic detection rate vs. standard surveillance testing (6 Aug-30 Oct 2020)
up to 4x
Asymptomatic detection rate vs. standard surveillance testing, peak season (August-September 2020)
85% more tests
Additional tests random testing would need to match Eva's detection
17%
Share of arriving households directed to PCR testing
>40,000 households/day
Households processed per day at peak season
40+
National entry points covered
Details
Maturity
Discontinued
Promoter
Greek Ministry of Civil Protection / Ministry of Tourism
Period
2020
Keywords
border management, public health, reinforcement learning, epidemiology, targeted testing, COVID-19
Context
Eva was a real-time reinforcement-learning system built as pro-bono work for the Greek government by an interdisciplinary team of operations researchers, epidemiologists and software engineers from USC Marshall, Wharton and AgentRisk. It was deployed across all of Greece's 40+ national entry points (airports, land crossings, seaports) from 6 August to 30 October 2020, during the COVID-19 pandemic.
Objectives
To allocate a limited daily COVID-19 testing budget among arriving travellers so as to maximise detection of asymptomatic infections, using anonymised traveller data submitted before arrival.
Activities
Travellers submitted a Passenger Locator Form at least 24 hours before arrival. Eva aggregated anonymised demographic and origin data, without retaining individual identifiers, to dynamically direct approximately 17% of arriving households to PCR testing, with the remainder receiving routine sampling. The Greek government shared real-time prevalence estimates with the EU to inform regional travel protocols, and the underlying algorithm was later open-sourced for international adaptation.
Results
The peer-reviewed Nature study (September 2021) found Eva detected twice as many asymptomatic infected individuals as standard surveillance testing, rising to up to four times during peak season (August-September); matching that detection through random testing would have required 85% more tests. The system processed data from more than 40,000 households per day at peak season and won the INFORMS Daniel H. Wagner Prize for Excellence in Operations Research Practice.
Conclusions
Eva is a rare technically well-evidenced example of AI-assisted crisis response, but the Hellenic Informatics Association raised governance and privacy questions about the system in November 2020 that were not fully publicly addressed, illustrating that strong technical evidence does not by itself resolve accountability concerns.
Implementation
Indicative cost
Medium (€50k–€500k) — Core system was developed pro bono by an academic/industry team; the practice description does not itemise the government's own operating costs for running the system across 40+ entry points.
Time to results
Short (< 1 year) — Deployed across a single COVID-19 tourist season, 6 August to 30 October 2020 (about three months).
Staffing & skills
Interdisciplinary pro-bono team of operations researchers, epidemiologists and software engineers (USC Marshall, Wharton, AgentRisk), Greek Ministry of Civil Protection / Ministry of Tourism officials administering border testing
Conditions for success
Mandatory Passenger Locator Form submitted at least 24 hours before arrival
Centralised, anonymised aggregation of traveller demographic/origin data feeding a dynamic daily testing-budget allocation
Coordination across all national entry points and real-time data sharing with EU authorities
Common failure modes
Hellenic Informatics Association raised governance and privacy questions about the system's conception in November 2020 that were not fully publicly addressed
Where it fits
Governance type
national government (Ministry of Civil Protection / Ministry of Tourism)
Scale
national, 40+ border points
Income level
high-income
Data sources
Where this practice's information was retrieved from, and when.
Statistics Sierra Leone and the GRID3 partnership use machine-learning building-footprint detection on satellite imagery to map population at 1-hectare resolution, …
