ABMs serve as a pivotal tool for unraveling hypothetical outcomes in complex systems. Here, we investigated the specific and joint effect of DCT and other NPIs on pandemic scenarios, which are governed by contacts between people. A component-based model architecture facilitates our approach, which encapsulates individually validated, realistic properties. The architecture builds on 3 model layers: behavior, pathogen, and NPI. To delve deeper into our model’s framework, encompassing its entities, parameters, and initial states, we guide readers to the ABM overview, design concepts, and details (ODD) [28,29] protocol (Multimedia Appendix 1) and Multimedia Appendix 2. Specifics of the modeling and simulation platform can be found in our Python-based implementation, which is publicly accessible on GitHub [30]. Figure 1 provides an overview of the Zeitgeber architecture, pathogen model, and examples of the virtual world infrastructure.

Behavior drives agents to engage in various interactions within their social networks of family, friends, and acquaintances, and experience encounters outside these circles, thus providing additional risk of contracting or transmitting the virus. In greater detail, we simulate agents’ activity patterns, including the nature and duration of activities, considering variables such as daytime, the agent’s current location, infection state (viral load level and infectiousness level). Therefore, we conceptualize NPIs as modifiers of the individual behavior that drives agent interaction, with the goal of reducing pathogen transmission. In other words, the NPI regime shapes the behavior of infected agents, their contact persons, and the population as a whole.

Relaxing NPIs allow agents to resume routine behavior, reducing the amount of time in confinement. Relaxing NPIs can reduce the social and economic burden of NPIs, for example, on public services such as gastronomy.

Under the RNT intervention, confined agents are required to obtain a negative test result before returning to normal activities before the default 14-day delay. We simulate a waiting period between test taking by sampling a distribution Y∼2+Γ(6,0.45). The wait period determines the duration in days between the last positive test and requesting a new test. After requesting the test, agents need to wait 24 hours for the test results. Without RNT, the standard recommendation for confined agents is to wait 14 days before resuming normal activities.

When an agent tests positive, they enter isolation and may notify agents in their contact network, thus triggering contacts to quarantine. During confinement, that is, isolation and quarantine, scheduling models that dictate an agent’s location changes are suspended. Agents resume their regular routines based on specific quarantine and isolation exit protocols.

During the COVID-19 pandemic, lockdowns and the emphasis on isolating and quarantining at home led to households becoming primary infection locations [65,66]. Our quarantine contact person household (QCH) intervention explores the potential impact of quarantining entire households when a member is isolated or quarantined. Thus, QCH aims to contain the virus spread within a single household, preventing further transmission to other households. QCH is implemented by using the Zeitgeber to stop allocating location changes for all agents in the household.

Closure of social facilities, for example, bars and restaurants, was prevalent during the pandemic [67-69]. To implement closure of bars and restaurants (CBR), we disabled the scheduling model that facilitates agents’ night-out activities. We focused on a static scenario, that is, the status of gastronomy services remains constant throughout the simulation run, thus reflecting the continuous closure of these establishments during certain periods of the pandemic.

Agents go into quarantine after receiving a contact notification through any contact tracing strategy and QCH, if applicable. The agent exits quarantine if one of the following conditions is met: (1) after the regular 14 days quarantine period has passed, (2) after receiving a negative test if RNT is used, (3) after all household members test negative, if QCH and RNT are applied together, and (4) by becoming infected, in which case, the agent is considered isolated and no longer quarantined. Recovered agents do not quarantine or comply with QCH after receiving a notification.

Symptomatic agents wait 1.5 days after symptom onset to get tested. Tests can be performed only during the test center opening hours between 8 AM and 4:30 PM. Agents receive test results after 24 hours. After receiving a positive test result, the agent may go into isolation.

Our calibration and validation efforts focused on selecting parameter configurations that let the ABM model respond in agreement with existing literature. We validated that the ABM model was able to reproduce activity patterns, contact patterns, and epidemiological observations specific to COVID-19. We performed sensitivity analyses for adherence, adoption, and compliance behavior parameters as they are the focus of this work. For simplicity, we marked the model outcomes as individual-to-mass behaviour (I2MB). Please see the Supplementary material in Multimedia Appendix 2 for a detailed explanation of the calibration and validation process.

Throughout this work, we use 95% CI using 100,000 bootstrap resamples to evaluate statistical significance since some of the quantities did not fulfill the normality assumption required to estimate the P value.

We compared daily instances per agent and daily cumulative activity duration per agent with the Extrasensory [70] dataset. The Extrasensory dataset was selected because it matched the activity granularity from our model and contained dense labeling for 1 week of study duration. We simulated 50 independent runs of 1000 agents; each run covered 7 days, without a pathogen model. We created an associative mapping of Extrasensory’s activities to I2MB activities. For example, lying down (not sleeping), watching television, and sitting at home labels in Extrasensory were mapped to OtherHomeActivities activity in I2MB. Please refer to the supplementary material in Multimedia Appendix 2 to obtain the complete mapping.

Our validation results show that CommuteCar, Eat, Grooming, OtherHomeActivities can be considered equal in the daily instances normalized per performing individuals, as the difference of means was not statistically significant. All other activities presented statistically significant differences of means between Extrasensory and I2BM, but we considered them not practically relevant as such differences did not exceed 2 instances per day. For daily activity instances normalized by total population, we observed that all activities but EatOut (CI −0.03 to 0.0) presented statistically significant differences of means. OtherHomeActivities (CI 3.51-3.96) and CoffeeBreak (CI 1.98-2.11) were the only activities where the difference of means exceeded the 2 instances per day. Nevertheless, we consider the differences, within a reasonable variation of expected habits, between the Extrasensory study population that primarily covered student campus life and our simulation that reflects a working population.

Daily activity durations show practical levels of agreement too (refer to Table S2 in the supplementary material in Multimedia Appendix 2). Distribution differences can be explained by the Extrasensory study population and its geographical location, that is, Extrasensory included university students from 1 city (San Diego, California). In particular, total daily commuting durations were shorter in Extrasensory, with a median of 20 min, whereas our simulated commutes had a median duration of 70 min in public transport (CI 46-54). In addition, OtherHomeActivity was consistently longer (CI 205-275), which can be attributed to the limited leisure activities provided in the simulation scenario. Detailed information about the validation of activity instances and duration is provided in the supplementary material in Multimedia Appendix 2.

We validated the contact patterns generated by the I2MB simulator against Tomori et al [71], who investigated contact person frequencies in different pandemic studies, referred to as waves. We replicated behavior patterns of POLYMOD and COVIMOD study conditions [71] by simulating (1) no intervention, (2) CBR, and (3) a population fraction staying at home.

Here we focus on the behavior of contact persons per location and overall. For simplicity, age and gender were not included in the I2MB simulator model. Unlike Tomori et al [71], we did not apply any normalization weights per household size as the simulations followed Germany’s household size distribution [72]. NPI restrictions during each COVIMOD wave were extracted from the first, fourth, fifth, and sixth Bavarian Infection Protection Measures Ordinance [73]. Please see the supplementary material in Multimedia Appendix 2 on how the waves were simulated.

Overall, our simulations matched contact person distributions for all study conditions. While there were statistically significant differences between the means, these differences were less than 2 contacts with the exception of others in POLYMOD (eg, 3.3 weighted contacts CI 3.23-3.40) and POLYMOD Work (eg, 15 weighted and with group contacts CI 14.41-15.59) where I2MB reported significantly lower contacts than the baseline. The latter difference can be explained by the group contacts assessment, where store workers would report their number of contacts in bulk. For simplicity, the special case of group contact assessment was not covered in our simulation.

The largest differences were observed for COVIMOD waves 1 and 2, which can be attributed to differences in average household sizes. For our simulations, the average household size was 2 residents. In contrast, the sampled population in POLYMOD and COVIMOD studies had an average household size of 3. Consequently, simulations showed 1 contact person on average, while COVIMOD reported 2. As restrictions were relaxed in subsequent waves, contact persons increased and nonhousehold contact persons became more relevant. We consider that the virtual world infrastructure always remains a limited abstraction of the real world, even if further spaces would be added. Thus, behavior in a modeled virtual world is already restricted. If NPIs do not show benefits in the modeled virtual world, they will likely fail in the real world too.

To evaluate the pathogen propagation model, we simulated the individual behavior of 1000 agents for a full pandemic duration. MCT and ICT were included in the simulations to align with the available real-world data reported in the literature.

Our pathogen propagation model has 3 functional parameters: infectiousness level, exposure function, and viral load removal function. Most model parameters were based on available evidence, for example, Kissler et al [53]. We empirically determined the viral load exhaled by an infected individual j as 16⋅ilj(t). Please see the supplementary material in Multimedia Appendix 2 and the ODD protocol for additional information in Multimedia Appendix 1.

While the emerging distributions for serial interval (CI −0.68 to −0.54 d), generation interval (CI −0.11 to −0.01 d), incubation period (CI −0.19 to −0.12 d), and illness duration (CI −0.47 to −0.40 d) had all statistically significant differences of means, the fraction of a day differences were practically irrelevant.

In addition to the epidemiological variables, we compared our simulated location of infection against the 30% at home-37% at work-33% at community rule [74]. We defined the infection location as the place where the exposed agent received the last load to pass the viral load threshold. Our analysis showed that ICT was necessary to match the rule at 29% (CI 26%-31%), 37% (CI 33%-40%), and 34% (CI 29%-41%) for home, office, and community. Without ICT, 64% (CI 64%-65%) of the infections happened at work and 34% (CI 33%-34%) at home, which corresponds to where individuals spent most of their time.

The behavioral approach used in this study yields contact networks with similar shapes to those generated by agent-based simulators [23]. However, the emergent contact networks in our simulation have contact-specific weights. To sensibly configure weights in a top-down approach that is typically applied in ABM simulators would necessitate data that may not be readily accessible. Figure S10 (supplementary material in Multimedia Appendix 2) illustrates the heterogeneity of networks per contact type, along with the corresponding weights indicating cumulative contact duration. Additionally, we calculated the median daily ratio of risk contacts at home vs all risk contacts (supplementary material in Multimedia Appendix 2 for details).

The behavior-driven ABM was implemented by the I2MB simulator [30] written in Python (Python Software Foundation). Our simulation generates and tracks the physical positions of 1000 agents in an appropriately sized virtual world, containing 500 apartments, 20 offices, 5 busses for public transport, as well as a hospital, a restaurant, and a bar (Supplementary material Figure S1 in Multimedia Appendix 2 and the Supplementary ODD protocol for further details in Multimedia Appendix 1). Simulations were performed with a time step of 5 minutes. Each pandemic simulation ran until there were no infectious agents in the population left (typically after 40-100 simulation days). Simulations were repeated 50 times for each parameter configuration, that is, specific behavior parameter settings, and NPI combinations; each simulation run had new agents and thus new home and office assignments, friends and acquaintances networks, and agent infectiousness profiles. At the beginning of each simulation run, 5 randomly selected agents were infected. We defined the default intervention as all symptomatic agents will be tested, receive a positive test result, and isolate upon the positive test. The agents wait for 1.5 days after symptom onset and request the test between 8 AM and 4:30 PM. The test takes 24 hours to complete. After 14 days, the positively tested agent can leave isolation and resume normal behavior. In addition, we applied MCT and ICT in all simulations.

As a consequence of our multishot viral model, when computing serial and generation intervals, it is not reasonable to assign the source of infection to a single agent. Therefore, we derived serial and generation intervals by considering all contact persons that contributed to an exposed agent crossing the viral load threshold and becoming infected. As a result, an infected agent contributes samples of serial and generation intervals, 1 for each considered contact person at the time of infection.

Our evaluation strategy first explored how behavior parameters (ie, adoption, adherence, and compliance) modulate DCT’s ability to control the pandemic. Subsequently, we explore how NPI combinations affect key pandemic characteristics.

To explore DCT effects on a population during a pandemic, we simulated parameter configurations for DCT adoption, adherence, and DCT compliance [12]. A total of 38,400 simulation runs were performed for the behavior parameter exploration (8×8 behavior parameter grid points, 50 simulation runs per grid point, 4 pandemic characteristics, and 3 constant behavior parameter set points). We included MCT and ICT methods for all simulations, as they would coexist in a real scenario (MCT and ICT set to default parameters; see Methods section). Guided by estimations for the COVID-19 pandemic, we selected the following set points for behavior parameters: adoption: 0.5 [75], adherence: 0.7 [76], compliance: 0.7 [77,78]. Furthermore, for each set point and associated grid, we uniformly sampled 1000 paths of length 3. Each step in the path varied only in 1 parameter. We used the sampled paths in a Morris [79] sensitivity analysis.

We evaluated NPI interactions by looking into 2 scenarios. First, we consider the optimal behavior scenario where adoption, adherence, and DCT and MCT compliance were set to 1. ICT compliance was left at the predefined compliance values (refer to Methods section for details). A realistic scenario was defined by setting adoption: 0.5 [75], adherence: 0.7 [76], and compliance: 0.7 [77,78] to estimated parameters from the COVID-19 pandemic. For the optimal scenario, we simulated pandemic progression for MCT and ICT alone, and for DCT in different combinations with RNT, QCH, and CBR. For the realistic scenario, we evaluated all possible NPI combinations. Thus, for 16 NPI combinations of realistic behavior and 9 NPI combinations of optimal behavior, a total of 1250 individual pandemic simulation runs were performed. Table 2 summarizes the scenarios and the behavior parameter settings.

To analyze DCT effects, we present a prospective simulation-based analysis that builds on contacts emerging from an individual, behavior-based modeling approach. Our ABM model uses parameters that had to be set empirically or informed by literature. The validation of key components (ie, population activities and behavior, agent contacts, and pathogen propagation properties) is detailed in Calibration and Validation section (in addition, see the Supplementary ODD protocol in Multimedia Appendix 1). We performed a parametric search across the DCT adoption-adherence-compliance space and evaluated the parameters’ combined effect on key pandemic characteristics. Moreover, we show the effect of selected NPIs in combination with DCT in optimal and realistic behavior scenarios.

Figure 5 visualizes simulation results for pandemic characteristics when sweeping behavior parameters (adoption, adherence, and compliance) specifically for DCT. Table 3 shows the results for the sensitivity analysis based on the parameter grids of Figure 5. We derived that any path toward the global maxima across all behavior parameters will unlikely yield a monotonic improvement of the pandemic characteristics. From Figure 5, it becomes clear that the public perception of DCT effectiveness may be discouraged during its implementation, thus further inhibiting DCT adoption. In more detail, the Morris sensitivity analysis (Table 3) suggests the following relevance ranking (decreasing order on the effect size μ*) across all pandemic characteristics, except for “total infected”: compliance, adoption, and adherence. For “total infected,” the ranking is adoption, compliance, and adherence, which is sensible because if people do not download and activate the app, they cannot comply. The parameter exploration furthermore suggests that at low adoption, the 10% region is not reached for any pandemic characteristic. Therefore, DCT adoption (ie, user base) should be maximized first in a DCT implementation campaign. Based on the fluctuation and corresponding sensitivity effect of “total infections” and “days in quarantine” across the behavior parameters, we conclude that adherence (25.2%, CI 25%-26% and 12.9, CI 12.6-13.2 d at 70% compliance) and DCT compliance (26.9%, CI 26.4%-27.4% and 13.8, CI 13.5-14.1 d at 70% adoption) are similarly important. However, after maximizing adoption, our results indicate that prioritizing adherence over DCT compliance may lead to the lowest 10% region of the pandemic characteristics, even at a DCT adoption far below the maximum. In more detail, the total infections and the 7-day incidence metrics approach a global minimum (low 10% boundary region), with DCT compliance at 70%, adoption above 90%, and adherence at 50%. In contrast, fixing DCT adherence at 70% requires adoption above 70% to access the lowest 10% region of each pandemic characteristic.

For all pandemic characteristics, we report the median of all the 50 runs with their corresponding IQR and 95% CI for difference of medians. Cohen d compatible effect sizes are reported in the supplementary material in Multimedia Appendix 2. The analysis shows that DCT, under optimal behavior, can improve pandemic characteristics compared with MCT and ICT alone. For example, without DCT, 65% (IQR 58%-69%) of the population are infected, whereas with DCT, infections reduce to 4% (IQR 3%-7%) at less than half of the wave duration (36 d, IQR 31-42 d). Moreover, days in quarantine are less with DCT than without it (−11 d, CI −12d to −10 d). For the optimal behavior scenario, additional NPIs, with the exception of RNT, do not show an additional effect on the pandemic characteristics. RNT in addition to DCT can reduce the days in quarantine by 10 (CI −10 to −9) days when compared with DCT alone. In summary, DCT has a profound effect on pandemic characteristics under optimal behavior, which could be improved only in combination with RNT if the goal was to reduce the days in quarantine.

For the realistic behavior scenario, Figure 6 illustrates pandemic characteristic ranges for different NPIs when applied with or without DCT. RNT has a negative effect on pandemic characteristics as it sends agents back to the vulnerable pool. Our results confirm that RNT, without DCT, could perform worse than MCT alone, for example, increasing “total infected” by 5% (CI 0.7%-10%). When RNT is applied in addition to DCT, “total infected,” “7-day incidences,” and “wave duration” increase and “days in quarantine” is reduced by 17 (CI −20 to −5) days with respect to DCT without RNT.

Using DCT in the optimal behavior scenario lowers the serial (5 d, IQR 3-7 d) and generation (5 d, IQR 3-7 d) intervals to range at or below the median incubation time of approximately 4.5 days. Hence, most interactions that lead to an infection occur at the end of the incubation period. In contrast, in the realistic behavior scenarios, “serial interval” (−1 d, CI −2 to 1 d) and “generation interval” (−1 d, CI −2 to 0.3 d) remain unaffected by DCT. The 30% of noncompliant agents in our realistic behavior scenarios with DCT applied, infected more people after their symptom onset compared with compliant agents. In particular, noncompliant agents increased “generation interval” and “serial interval” to 5 (IQR 3-7) days each, compared with 4.5 (IQR 3-6) days and 4.7 (IQR 3-7) days in the optimal scenario. Due to noncompliant agents, generation and serial intervals are similar to the configuration without DCT. While numeric results are calculated using medians of means per run, we randomly selected 50 agents with “days in quarantine,” and 50 instances of “generation interval” and “serial interval” each, measured from infectious-infected agent pairs, to build the corresponding distributions in Figure 6.

Different NPI combinations may yield similar results when applied together with DCT, as illustrated in Figure 6. By “total infected,” DCT alone (38%, IQR 26%-45%) performs similarly to DCT with QCH and CBR (38.6%, IQR 20%-47%), for example, a not statistically significant difference of 1.6% (CI −21% to 12%). Furthermore, DCT and QCH (30%, IQR 18%-41%) perform similarly to DCT and CBR (32.6%, IQR 11%-43%), for example, no statistically significant difference of 8.7% (CI −31% to 19%). The best pandemic characteristics are found for applying DCT and either QCH or CBR, but not both. Further interpretation is provided in the Discussion section. For a detailed comparison between NPI configuration, refer to the supplementary material in Multimedia Appendix 2.

A side effect of DCT is that the median false discovery rate increases from 88.4% to up to 97% for optimal behavior. When introducing DCT in realistic behavior scenarios without RNT, the false discovery rate is approximately 93%, depending on the combination of QCH and CBR. In contrast, the false discovery rate for realistic behavior scenarios drops to approximately 76% when using RNT in combination with DCT, and to approximately 62% without DCT. Without DCT and RCT, combinations of QCH and CBR have a median false discovery rate of approximately 88% (supplementary material in Multimedia Appendix 2).

In our simulation of realistic behavior scenarios, the false negative rate was between 26% and 32%, and for optimal scenarios, the false negative rate varied between 24% and 36%. False discovery and false negative rates are shown in Table S9 in Multimedia Appendix 2 for the optimal and in Table S10 in Multimedia Appendix 2 for realistic behavior.

We note that when applying RNT, the false negative rate was larger for the optimal than the realistic scenarios including DCT. By analyzing the optimal scenario, we found that false negatives are generated by (1) unnotified asymptomatic agents, and (2) differences between our agent pathogen model and the standard DCT app behavior. Due to the asymptomatic proportion of 40% (n=1000), it is likely that infection chains of exclusively asymptomatic agents occur, which remain undetected by DCT. As RNT lets susceptible agents return to normal activities, they may get in contact with asymptomatic agents and hence get infected. In contrast, in a realistic scenario using DCT, noncompliant agents, while being infectious, continue to trace their contacts. Thus, upon the contact between a noncompliant and an asymptomatic agent, DCT notifies the latter of an at-risk contact, as well as recommends testing and subsequent isolation of the asymptomatic agent.

Another reason for false negatives, even in the optimal behavior scenario, is a discrepancy between our multishot [44,51] pathogen model and the simulated DCT app behavior. According to our pathogen model, agents who accumulate sufficient viral load through multiple exposures of less than 15 minutes each will eventually get infected. However, our simulated DCT app follows the original World Health Organization (WHO) recommendations [80], where infection is assumed after 15 minutes of continuous exposure (ie, one-shot assumption [44]), thus effectively failing to capture agents that got infected through multiple shorter exposures.

We investigated the impact of individual behavior on the effectiveness of DCT and the interaction of DCT with multiple other NPIs to guide future pandemic control efforts and prioritization for pandemic preparedness. Our methodology is motivated by the observation that NPIs can be encoded as behavior rules that affect individual choices and actions. Consequently, a microlevel behavior-driven ABM simulation was proposed, where a Zeitgeber architecture orchestrates activity patterns for every agent that lead to varied agent interactions and key behavior that affects pandemic progression, for example, staying home. Along with the detailed, behavior-driven ABM, we deployed a deterministic multishot viral load transfer model for respiratory pathogens. Our behavior-driven ABM was able to generate realistic pandemic characteristics, reproduce complex phenomena, including stochastic extinctions (Figure 6), and highlight the complex interaction of NPI combinations. Furthermore, results show that individual behavior in general, and related to DCT (adoption, adherence, and compliance) in particular, affects DCT’s capacity to control a pandemic. Our findings suggest that, when composing NPIs, more is not always better; the performance of NPIs (eg, to reduce or delay infections) varies depending on how the interventions are composed (Figure 6).

Our analysis demonstrates that DCT alone profoundly reduces pandemic impact if DCT-related behavior is optimal, with a 92% reduction of the total infections. In realistic behavior scenarios, DCT still reduces total infections by one-third compared with only MCT and ICT. Furthermore, our results show a substantial synergy between specific NPI combinations. In particular, combining DCT with CBR or QCH in realistic behavior scenarios yields the best overall performance; that is, reducing total infections by 20% (CI −36% to −4%), from 38% (IQR 26%-45%) for DCT only to 30% (IQR 18%-41%) for DCT and QCH. Retrospective analyses [7] on the importance of combining DCT with other NPIs corroborate our findings. However, even if individual NPIs have a positive effect on pandemic characteristics, their combination may not, as observed in our results when applying both QCH and CBR with DCT, which is statistically similar to only using DCT with a reduction of only 2% (CI −20% to 11%). As individuals spend more time in close proximity at 1 location, they may form a locally dense contact network and thus infect others. Here, QCH and CBR may have forced more agents to spend time at the home location with an infected household member. However, the reduction in the median daily ratio of risk contacts at home to all risk contacts for QCH vs QCH and CBR is less than 0.1% (supplementary material in Multimedia Appendix 2), suggesting complex interaction mechanisms between NPIs. Further investigation with a behavior-driven modeling approach is needed to understand NPI combination effects.

We found that DCT increases the false discovery rate; that is, DCT leads to a higher proportion of individuals being quarantined. Conversely, DCT decreases the false negative rate; thus, fewer infected individuals are left undetected and unisolated. Moreover, our results show that DCT confines more people per risk contact than any other NPI considered in this analysis. Due to the comparably large false discovery rate of DCT, more of the asymptomatic individuals get notified; thus, tested, and subsequently isolated than without DCT.

Although DCT adoption plays an essential role in pandemic control [4], our analysis confirms that maximum DCT adoption is unnecessary to reduce key pandemic characteristics [81,82]. For example, an adoption of 50% in the realistic behavior scenario yielded a reduction of the total infections by 42%. While DCT introduction campaigns should aim at improving adoption and adherence, pandemic characteristics do not react homogeneously (Figure 5), which may negatively influence public perception. Therefore, it is important to maintain ongoing public motivation strategies to offset any short-term setbacks. Additionally, it is not appropriate to judge the effectiveness of DCT solely based on short-term pandemic indicators, as they might give a misleading impression of DCT’s ineffectiveness in pandemic management.

In a typical ABM, the predefined social networks are characterized by their node degree; that is, the contact person count per agent. Predefined networks are often static; that is, not evolving over time. However, behavior changes induced by NPIs change networks. The relationship between behavior changes and network dynamics is often neither observed nor derived. Consequently, often unverified assumptions need to be made about how behavior changes affect the networks. With a behavior-driven ABM, we replace assumptions about how behavior changes affect network dynamics with assumptions about pandemic-independent, verifiable behavior patterns that can be altered by NPIs. Accordingly, agent interactions create the network dynamics. Emergent networks allow us to evaluate metrics, including degree (ie, contact person count) and strength (ie, aggregated encounter duration across all contacts). Further upscaling of the approach should focus on adequate interaction environments, rather than just increasing population. For example, upscaling the population in the current work would likely maintain the statistical relationship of home and office networks as reported, and thus maintain the match with real data (supplementary material in Multimedia Appendix 2). Network diversity could evolve from adding additional public spaces, for example, parks and shopping malls, where super-spreader events may occur.

Our analysis of network metrics and pandemic characteristics revealed epidemiological implications. In particular, our network dynamics analysis (Supplementary material in Multimedia Appendix 2) showed that infection time and 7-day incidence correlated with network strength rather than network degree. In addition, network strength reflects the change in contact preferences, if an NPI modifies individual behavior. Consequently, the reported risk increase for household members due to isolation [65] is not connected to the average degree of the family network. While the average degree of the family network did not change in our analysis, the average family network strength rose during isolations, which may explain the increased infection risk of household members during isolation periods.

DCT advances over MCT by removing recall issues associated with location, familiarity of contact persons, and time since last contact [61]. Furthermore, a DCT app could track short encounters and is not limited by workforce availability to collect and process contact information. While our settings for MCT focused on natural process delays and did not elicit capacity limitations for the considered population size, the addition of DCT clearly improved pandemic characteristics (Figure 6). Based on our results, we conclude that DCT outperforms MCT risk assessment not only due to unrestricted capacity, but the reduction of contact processing time.

During the COVID-19 pandemic, the US Centers for Disease Control and Prevention (CDC) changed the initial contact tracing rule from 15 minutes of continuous exposure to a cumulative exposure duration of at least 15 minutes within 24 hours [83], which corresponds to our multishot approach in the pathogen model. Our analysis of false negative cases confirms the CDC decision. More infectious pathogens may spread during shorter interactions. For example, at the emergence of the omicron variant, the US CDC further shortened the duration threshold to 10 minutes, due to the variant’s increased infectiousness [84]. At our current simulation time step of 5 minutes, the omicron variant spread could still be represented, but short exposure risk would be underestimated. The simulation time step could be reduced to represent short activities and their exposure risk. Besides the pathogen, DCT technology may affect the simulation time step. For instance, the Apple and Google Exposure Notification Framework [85] implemented a dynamic delay between scans in the range of 2-5 minutes. For investigations on the sampling mechanism, a simulation time step of less than 1 minute is justifiable. However, the simulation time step relates indirectly to the simulation time and processing costs. The simulation time step should be chosen carefully to balance simulation resolution with information gain.

The evaluation of guidelines for isolation and using rapid antigen tests identified that unnecessary quarantines produce an avoidable economic burden [86]. Therefore, days in quarantine can be considered a proxy for the economic burden of individuals for whom working from home is not an option. Our findings suggest that RNT is the only NPI that directly reduces the economic burden of quarantine. However, the quarantine reduction by RNT comes at the cost of increases in other pandemic indicators, including 7-day incidences, total infections, and wave duration (Figure 6).

Mobility-restricting NPIs are disregarded in public opinion. Thus, QCH might be more challenging to implement than CBR. However, our findings indicate that CBR and QCH offer limited differential impact on pandemic characteristics. Therefore, the economic benefit of keeping facilities operational, for example, restaurants, might outweigh the costs associated with enforcing QCH.

In our modeling approach, behavior and pathogen transmission are modeled independently. The separation allows modelers to distill complex behaviors into physical interaction parameters, that is, distance and duration, which are then used to control viral load transmission. However, modeling viral load transfer and resulting infection needs to align with behavior modeling. As activities and interactions become more fine-grained, the quantization used to describe virus load transmission must follow. The versatility provided by our Zeitgeber architecture and microlevel, behavior-driven ABM approach pairs with our proposed multishot pathogen model. As a result, we can model the composition of different NPIs and implement regulatory changes like the definition of “close contact” by the US CDC (refer to Effectiveness of DCT and Its Combinations With Other NPIs section).

It is recognized that human behavior plays a critical role in pandemic development, from individual decisions to the level of policies, for example, NPIs. Work by Wang et al [24] and Blanco et al [87] presented simulations of disease spread, including NPIs. While we share their objective to achieve realistic simulations of everyday behavior within detailed small-scale cityscapes, including, for instance, offices and restaurants with typical environmental features and behaviors, our work advances in several key areas. Our approach follows important epidemiological principles by incorporating a pathogen model with multishot viral load transfers. Moreover, we analyze key epidemiological metrics (Table 1). Blanco et al [87] asserted the potential of crowd simulation, in particular at the “atomic” level of individual agents, where infection occurs. The analysis in this study is explicit about emerging macro- and meso-level quantities (eg, total infections and confinements stratified by notification type, see video in Supplementary material in Multimedia Appendix 2). In summary, our work breaks new ground in behavior-driven modeling and simulation by integrating rigorous epidemiological principles and ABM methodologies, thereby offering a robust and scientifically grounded approach for analyzing disease spread and assessing public health interventions.

While the selected pathogen characteristics are well-established, for example, the serial interval, there is no consensus on how to evaluate the combined effect of multiple NPIs. Here, we selected pandemic metrics that are generally recognized and used during the COVID-19 pandemic. Consequently, we did not emphasize metrics that apply to DCT only, for example, turn-around times [12]. Several pandemic metrics (7-d incidence metrics, days in quarantine, and total infections) were derived due to their relevance for different stakeholders, including individuals, the health system, and the public.

Our approach could be extended to design and test NPIs that have not been applied before. In particular, NPIs that depend on a person’s characteristics, for example, household size and viral load profile, may be analyzed with our approach and, more specifically, by reconfiguring the platform provided with this work. The novel NPIs and their combinations would be comparably more challenging to describe in compartmental models, which assume that agents are interchangeable.

Our ABM modeling approach improves over existing models in the following ways. First, we break the coupling between behavior, NPI rules, and pathogen transmission, thus providing modelers with fine-grained control over their simulated population. In contrast, many existing models do not have a direct correspondence between parameters and individual behavior changes. For example, the compartment transmission rate in a susceptible, exposed, infected, and recovered (SEIR) model condenses many behavioral and pathological aspects into a single parameter that is hard to associate with behavior changes. Second, our ABM model does not rely on the observations of the pandemic evolution. Therefore, scenarios outside pandemic or epidemic events could be studied, catering to pandemic preparedness. Third, our approach supports context information associated with individuals and pandemic control, which may allow policymakers to consider the effects of geographical, structural, and socioeconomic dynamics.

It is clear that modeling a virtual world will only approximate actual behavior and facilities (Figure 1). For a behavior-driven ABM, virtual world complexity, daily routines, location detail, activity diversity corresponding to routines and locations, as well as agent count, are main factors that affect simulation fidelity and generalizability. Nevertheless, a behavior-driven ABM simulation with the multishot viral load transmission mechanism could serve as a first exploration step when designing and orchestrating NPIs. As if an NPI is ineffective in the constrained behavior-driven ABM simulation, it is unlikely that the NPI will create benefits in reality. For example, while we investigated perfect behavior scenarios, the NPI combinations showed limitations, with results indicating that about 3.5% (1000) of the population was still getting infected. The simulation should therefore be considered as an expected performance boundary, and used to select promising NPI combinations.

For SARS-CoV-2, the original pathogen transmission characteristics were understood within 6 months after the initial outbreak [54] and subsequent variants were characterized within weeks [88,89]. NPI effects on pandemic characteristics require additional time in extent of weeks to months to be determined in an ongoing pandemic [90]. However, quickly understanding the effects of NPIs is essential for policymakers to control pandemic development [91]. Moreover, the population may perceive an NPI and potential NPI adjustments as unnecessary grievances as they restrict personal freedom [67]. Furthermore, evaluating a single NPI’s effect on pandemic characteristics is infeasible as effects cannot be isolated in reality. Behavioral simulation is therefore essential to design and test NPIs regarding their short-term and long-term effects and before implementing them in reality. Our approach equips policymakers with a tool to understand an NPI’s effectiveness for pandemic control, for example, in 7-day incidences and total infections, compared with collateral damage expressed in wave duration, days in quarantine, days in isolation, false discovery rate, and false negative rate. For example, policymakers can use the simulation to analyze the benefits of implementing RNT as a measure to reduce economic burden.

Since simulated agents are not required to match actual individuals, no privacy-critical data from the citizens are required to evaluate the NPI’s effectiveness. With behavior-driven ABMs, no data from previous NPI implementations are necessary to create and simulate contact networks. Consequently, in a virtual world representation and assuming that the underlying pathogen transmission characteristics are known, the NPI analysis time depends on simulation capacity only.

To investigate behavior effects related to DCT and NPIs, we kept ABM model parametrization static for several components, for example, the viral transfer model. Therefore, our analysis does not provide conclusions across wider viral characteristics, location arrangements, etc. However, our current findings provide meaningful insight on NPI composition performance for contact tracing in complex realistic behavior scenarios under a COVID-19 pathogen.

Our current analysis was focused on Western culture and daily repeated activity patterns in an urban setting, using office spaces, public transport, and bars and restaurants as a surrogate of informal interhousehold interaction. Future studies could explore additional activity patterns influenced by culture, age, and work vs leisure days. Furthermore, other interhousehold activities, including visits to other agents’ homes, could be added to provide additional infection pathways. Moreover, the pathogen model could be adjusted for virus variations. While our approach can be extended to more diverse conditions as described above, our current results illustrate the potential, limitations, and contributions of NPIs, including DCT, alone and in concert. The model complexity [92] of a behavior-driven ABM is controlled by simulation population size n. While activities, behaviors, and locations create a linear complexity increase, interaction-based NPIs, for example, contact tracing, have a worst-case quadratic complexity, that is, O(n2). Our implementation averages a complexity of O(n1.3) (refer to Runtimes section in supplementary material in Multimedia Appendix 2), resulting in run times of 55 minutes for n=1000 and 21 hours for n=10,000. The simulation was timed on an Intel Core Ultra 7@4.8 GHz with 68 GB of RAM.

Future work may extend the modeling and simulation and consequently refine understanding of the underlying interaction mechanisms, while still managing complexity. We argue that the flexibility of behavior-driven ABM is 2-sided. While our approach offers countless ways to explore uncharted conditions, it burdens the modeler with decisions on fidelity, thus potentially adding structure without necessarily enhancing understanding and results.

Behavior-driven ABMs add important insight on DCT effectiveness as well as help to understand the impact of combining and fine-tuning NPI combinations that maximize effectiveness, minimize collateral damage, and avoid more disruptive interventions, including proactive school closures or general lockdowns. Compared with predefined network ABMs, behavior-driven ABMs are easier to adjust and to interrogate regarding specific environments, population routines, and virus properties, and provide additional transparency and explainability that helps to disentangle assumptions. For example, the complexity of behavior patterns inside locations can be modified independently of the pathogen transmission mechanics, which helps identify risky behaviors, rather than assuming a priori that certain behaviors are risky. Future investigations may model additional detail in the pathogen model parametrization, the virtual worlds, and the choice of metrics to measure social and economic impact. Based on the simulation results for a selected region of interest, policy design can balance pandemic impact with cultural differences across communities, thus preventing possibly discriminatory NPIs.