Exploring the effectiveness of a COVID-19 contact tracing app using an agent-based model

The study addresses whether a smartphone-based contact-tracing app (CTA) can effectively mitigate COVID-19 spread in a post-lockdown context and under what conditions of testing capacity, testing policy, and user behaviour it is most effective. With non-pharmaceutical interventions central to control efforts and manual contact tracing limited by delays and inefficiencies, Bluetooth-based digital contact tracing has been proposed. Prior analytical models suggest epidemic control with rapid tracing and high adoption (around 60%), though concerns remain about privacy, feasibility, and unintended consequences such as excessive quarantining or ineffectiveness under certain policies. The purpose is to simulate realistic social interactions and evaluate how CTA adoption, testing capacity, and compliance with self-isolation interact to influence epidemic dynamics, thereby informing policy on testing strategies alongside digital tracing.

Existing work indicates that instantaneous or rapid contact tracing can aid epidemic control, with some models positing that roughly 60% app adoption could suppress transmission. Other modelling highlights potential downsides, including large numbers of people isolating or limited effectiveness compared to alternative strategies like random testing. The literature also underscores heterogeneity in contact patterns as crucial for tracing effectiveness and the limitations of homogeneous compartmental models. This study builds on these insights by using an agent-based framework to capture heterogeneity and multi-layered contact structures, and by jointly examining adoption, testing capacity, testing prioritization policies, and behaviour such as compliance with isolation.

The authors developed an agent-based model (ABM) of COVID-19 transmission grounded in SEIR principles, simulating an urban population of approximately 103,000 agents representing Glasgow, UK, derived from the 2011 UK Census. Agents are heterogeneous (age, sex) and embedded in a multi-layered social network: households; non-cohabiting relatives; workplaces (with clustering of close colleagues and additional random coworker contacts, and 13% in customer-facing roles); friendship networks (Barabási–Albert scale-free, median 14 friends, age-assortative); school classes (ages 6–17, up to 30 pupils per class). Daily contacts are simulated by domain with specific frequencies and contact counts; elderly agents have reduced social contact frequency. Random and public-facing contacts are sampled from Poisson distributions proportional to local zone populations (zones ~200–2700 residents). Transmission probabilities differ by contact type, with lower risk for brief/random encounters; infection susceptibility for under-16s is reduced by 50%. Disease progression follows age- and gender-dependent probabilities and durations: incubation with potential pre-symptomatic infectiousness (25% chance to become infectious on each of the last three incubation days); branching to asymptomatic or symptomatic (mild or severe) states; declining infectiousness for asymptomatic after day 3; severe cases progress to hospitalization (no further contacts assumed). Infectious period after symptom onset is 7–11 days. Hospital capacity is assumed sufficient and no nosocomial transmission is modelled. Mitigations include a CTA for agents over 14, which records proximate CTA contacts over the previous 10 days and can notify contacts upon a positive test. Testing is modelled with fixed daily capacity and 1–3 day delay from symptom onset to test, with results within a day. Agents seek tests when symptomatic or when notified (via relatives’ alerts, classroom quarantine after a pupil’s positive test, or CTA notification). To reflect influenza-like illness (ILI), 3.5% of the population weekly experiences ILI; 30% of them seek testing and test negative, consuming capacity. Self-isolation behaviour removes all social ties except household; household transmission probability is reduced by 30% during isolation. Compliance is probabilistic: agents self-isolate when symptomatic with mean probability ω=70%; app-notified but asymptomatic agents have reduced isolation probability by factor Ω (0<Ω<1). The testing and isolation decision flow is specified (including priority policy vs first-come-first-served). Calibration: Baseline ‘business as usual’ contact patterns were validated against a UK contact survey, showing similar lognormal-like distributions (mean ~20, SD ~13 contacts/day). Transmission probabilities were calibrated to R0≈2.8 in the first three weeks (β parameters around 8% for network contacts and 0.8% for random). A post-lockdown baseline with social distancing assumed 3 days/week attendance at workplaces and schools, 30% reductions in school, random, and friendship contacts, and 30% reduction in non-household transmission probabilities, yielding mean ~14 contacts/day and R≈1.5. Experimental design: 140 scenarios (20 stochastic runs each) varied CTA adoption among over-14s (0, 20, 40, 60, 80%), weekly testing capacity (0, 0.5, 1, 1.5, 3, 6%, unlimited), CTA isolation compliance (high Ω=0.9 vs low Ω=0.5), and testing policy (priority to symptomatic vs no priority/first-come). Initial conditions reflected UK estimates at reopening: 7% recovered and 0.3% infected at start.

- Testing alone: Increasing testing capacity from 0 to 3% of the population per week reduced overall infections from 44% to 31% and lowered peak infections by 59%. Beyond 3%, additional capacity had little added benefit because peak daily prevalence was ~3% and about half of infections are symptomatic, allowing available tests to cover symptomatic cases plus contacts and some ILI demand. - CTA with priority testing of symptomatics: As CTA adoption increases, both cumulative infections and peak prevalence decline across testing capacities. With 80% CTA adoption and unlimited testing, cumulative infections fell from 45% to 15% and peak cases fell by 89%. For intermediate adoption (40–60%) and testing (1.5–3%/week), cumulative infections reached 22–27% and peak reductions were 70–85%. Even without routine testing, some benefit appears due to notifications from severe cases diagnosed in hospital. - CTA without priority for symptomatics: With limited testing and no priority, higher CTA adoption can increase total infections due to test stock depletion by mostly uninfected notified contacts, reducing the ability to test symptomatic cases. For example, at 3% weekly testing, 40–80% CTA adoption produced substantially more infections than no CTA. Efficiency (share of positive tests) is higher when prioritizing symptomatics for any adoption level. Unlimited testing eliminates this drawback and shows consistent benefit from CTA. - Compliance effects: High vs low CTA user compliance with isolation changes cumulative infections by only a few percentage points (~3–5%) but affects peak suppression: peak reduced by ~80% with high compliance versus ~74% with low compliance. Impacts of compliance are greater at higher adoption and lower testing capacities. - Overall, CTA effectiveness is enhanced by sufficient testing capacity and prioritization of symptomatic individuals; otherwise, increased demand from notifications can be counterproductive under constrained testing.

The findings indicate that app-based contact tracing can substantially mitigate COVID-19 spread within a social distancing context, provided testing capacity is adequate and testing policies prioritize symptomatic cases. CTA adoption generates two opposing effects: it facilitates earlier isolation of infected individuals (including asymptomatic/presymptomatic), but it also increases testing demand from many uninfected notified contacts. Prioritizing symptomatic individuals and expanding testing capacity shift the balance toward benefit, reducing both cumulative infections and peak burden. Policymakers should scale testing in line with CTA adoption and implement prioritization rules to maintain testing efficiency. Even with lower compliance among app-notified users, meaningful reductions are achieved, and compliance can be further supported through timely testing and socio-economic support for isolation. The synergy between social distancing, testing, and digital contact tracing helps flatten the epidemic curve and protect health system capacity, particularly during phases of higher viral circulation.

Smartphone-based contact tracing is a viable mitigation strategy that complements manual tracing, offering speed and potential cost efficiencies. As adoption increases, population-level benefits grow by breaking transmission chains, even though individual users may not directly benefit. Effective deployment requires pairing the app with sufficient testing capacity and prioritization of symptomatic cases. The study’s agent-based approach demonstrates substantial reductions in cumulative and peak infections under appropriate policies, supporting continued development and integration of digital tracing in public health responses.

The model relies on assumptions and parameters that reflect emerging evidence and may not capture all real-world complexities. Transmission and progression parameters were calibrated to match aggregate R0 estimates and validated against contact survey distributions, but alternative combinations of contact patterns and transmission rates could reproduce similar aggregate dynamics while altering the domain-specific contributions to spread. Sensitivity analyses indicate the qualitative findings about CTA impact are robust, but quantitative outcomes may vary. The model assumes homogeneous app performance without technical failures or notification delays, and accurate test results returned within one day. Demographic heterogeneity in app adoption is not modelled and could influence outcomes. Hospital capacity constraints and nosocomial transmission are not included.