Feasibility of digital contact tracing in low-income settings – pilot trial for a location-based DCT app

To assess the DCT app effectiveness, we assembled aCT hits and pCT hits of all participants designated as infected, with corresponding matches (Table 4). To evaluate the extent of tracking activity, Fig. 6 displays the percentage of tracking participants and the mean tHR per day. By pCT, 1075 hits of designated infected participants with other participants have been listed, resulting in a median of 25.8 (± 28.8) hits per day per designated infected participant. A total of 147 pCT entries were not analyzed due to incomplete documentation. With aCT, five hits with a median of 0.1 (± 0.6) hits per day per designated infected participant were recorded. Of ten designated infected participants, only four were recording GPS coordinates during the infected period, of whom one participant only recorded one GPS coordinate. All designated infected participants recorded 4.3% (42.4 h tHR) of the time they were infected (984 h mHR) and a median of 62.1 minutes per day per infected person. The average time all participants logged GPS coordinates was significantly lower (p < 0.001) at 15.1 minutes per day per person.

This provides an explanation for why only five hits were recorded via the DCT app compared to the possible 1075 pCT listed hits. Of those five aCT hits, two were also recorded via pCT and could be accounted for as matches. Two aCT hits were recorded via Bluetooth pairing while GPS was inactive; thus, the DCT app used prior logged GPS coordinates that no longer reflected current location and resulted in a false-positive hit. One aCT hit was not listed via pCT by the designated infected participant but by the contact person and thus could be identified as a true aCT hit. In this case, the hit was recorded via aCT but not via pCT, demonstrating that the DCT app can be superior to a paper-based system on the precondition that both participants use the DCT app correctly.

Table 5 shows the contingency table evaluating the effectiveness of the DCT app. If every hit was a true hit, pCT found 1075/1077 (99.8%), and aCT found 5/1077 (0.5%). If only hits were analyzed when both participants had switched on their GPS, all three pCT hits were also recorded via aCT (plus one true aCT hit that was not listed via pCT). This suggests a high sensitivity of the DCT app under efficacy conditions. Determination of reliable thresholds for sensitivity or specificity was not possible due to the reduced time of tracked movement (tHR).

To determine if time stamp differences between GPS and Bluetooth pairing logged coordinates (periods where GPS was switched off, but Bluetooth was switched on, and thus Bluetooth pairing did not reflect the current location) were a relevant problem, the movement profiles concerning this matter were analyzed. Of the 101 participants where GPS coordinates were sent to the server, 47 participants (46.5%) showed periods of Bluetooth pairing and switched off GPS between one to ten times during the study, with a median time of 3:33 (0:31–14:09) h. Considering that 46.5% of the DCT app users for 3:33 h had an augmented possibility to record false-positive hits, this was identified as a significant impairment under real-world conditions.

To further examine the performance of the DCT app, all days on which designated infected participants tracked their movement using GPS with no aCT hits being recorded were analyzed (Table 6). There was no pCT hit where both infected and regular participants tracked their movement using GPS but no aCT hit was recorded. Consequently, no false negative results for aCT were found.

To evaluate aCT accuracy in buildings, an index smartphone was positioned next to participants. To detect whether Bluetooth increases accuracy in buildings as suggested, a comparison was made between aCT hits based on coordinates logged by both GPS and Bluetooth pairing and those based on GPS alone. For 7/22 participants, the hit only was recorded using additional coordinates logged by Bluetooth pairing. The results shown in Fig. 7 demonstrate that in buildings, coordinates logged by Bluetooth pairing increase the accuracy of the DCT app. The mean distance recorded by the DCT app using GPS and Bluetooth was 22.9 m ± 21.6 SD and was significantly (p = 0.004) more accurate than 60.9 m ± 34.7 SD recorded only via GPS coordinates. No significant correlation between recorded distance and phone brand or version of the smartphone OS was found.

To estimate accuracy in public transport, participants were asked to track their route to work using the DCT app. The precision of every GPS coordinate to the track was measured and illustrated in a boxplot (Fig. 8). The mean distance between the idealized route and recorded GPS coordinate as an approximation for GPS accuracy is 10.3 m ± 10.05 SD. There was a significant (p = 0.007) correlation between precision and phone brand, but no correlation with smartphone OS was found.

To evaluate the accuracy outdoors in an urban environment, GPS coordinates of participants positioned at distinct locations were logged. The precision of logged GPS coordinates in relation to distinct location points was measured and illustrated in a boxplot (Fig. 9). The mean distance between the distinct location and recorded GPS coordinate as an approximation for GPS accuracy outdoors is 10.4 m ± 4.2 SD. No significant correlation between degree of precision and phone brand or the version of the smartphone OS was found.

Additional findings that could be relevant for the implementation of a DCT app in low-income settings are as follows:

The effectiveness of mobile app-based DCT in real-world conditions of low-income settings has been identified as a major challenge by this study. High variability in smartphone quality, low levels of adherence to apps, insufficient access to mobile internet and unreliable access to electricity supply all inhibit DCT effectiveness. Many of these findings are likely to be an issue in other low-income regions as well and should be considered by further initiatives implementing mobile health tools. Sachs et al. [2] reported similar findings while introducing a mobile health tool during the 2014–16 Ebola epidemic in West Africa. There are recent studies on the effectiveness of DCT apps in real-world settings in La Gomera, Spain [11], Switzerland [12], Norway [13] and the UK [14] that conclude positively about the performance of the app. The results are difficult to compare with our findings because the effectiveness was estimated via indirect parameters due to the decentralized approach of the apps. A prospective study exploring the DCT app used in Australia [15] used a centralized approach and thus could compare DCT and MCT hits. The reported findings on effectiveness were similar to our results, and they concluded that the app did not make a meaningful contribution to COVID-19 contact tracing. Explanations for why many countries struggle to exploit the potential of DCT as predicted by modeling studies [3, 16‐18] are the focus of current research. Based on a review of the literature [9], privacy concerns over user data, low trust in government and third parties (big data analysis, malicious actors), security vulnerabilities (hacker attacks), ethical issues (discrimination of minorities), user behavior and participation (limited experience with mobile devices, reasons for adoption) have been prioritized in primary studies, in contrast to understanding technical constraints. This is likely indicative of concerns around adoption of a DCT app dominating technical limitations in high income settings, which reflects the view of countries where primary studies have been conducted. According to our study results, technical constraints are a greater barrier to DCT implementation in low-income setting countries. Factors such as battery consumption in combination with limited energy supply, limited internet access (costs for mobile data, network coverage), hardware issues (smartphone quality, Bluetooth compatibility), and software issues (Android OS fragmentation, replicated software) should be the focus of further research.

The studied DCT app is a location-based approach to DCT, using GPS to calculate proximity to infected individuals, while most initiatives of DCT have been based on proximity inference via Bluetooth Low Energy (BTLE). Using BTLE over GPS would reduce battery consumption, partly mitigating unreliable energy supply effects on DCT effectiveness. Technical reliability limitations of BTLE have been reported [19, 20]. Our finding that newer versions of OS impede proper app use has been reported by other initiatives and was addressed by the new Apple/Google Application Programming Interface (API) [21]. However, BTLE is an emerging technique and cannot be applied on every smartphone, especially in low-income settings. Given that more than 40% (one billion) of Android active users worldwide use version 6.0 or below and no longer receive updates, many Android devices may not benefit from updates to the new BTLE-based contact tracing system Google built in collaboration with Apple during the COVID-19 pandemic (Google/Apple API) [10]. High app uptake rates have been identified as a crucial factor in DCT effectiveness [3, 17, 18], and unsupported OS versions could prevent achieving the required uptake rates in low-income settings if BTLE is used. Of the participants in our study (urban population, hospital staff), 20% were using Android 6 or below, whereas in the overall population, the number could be significantly higher. Thus, a large proportion of smartphone owners in low-income settings would be excluded from DCT if BTLE techniques are selected. Additionally, the existing difference in access to health care due to DCT for disadvantaged communities caused by limited availability of smartphones and internet access [22, 23] would be further exacerbated by implementing a DCT app requiring BTLE. In contrast, the location-based Bluetooth-assisted GPS method used in this study mitigates those barriers, as it utilizes widely available technology.

To address the problem that data transmission of the DCT app was often limited due to insufficient mobile data credit, we strived for a solution to provide mobile data for the DCT app free of charge. Approached ICT providers were willing and able to realize that in the event of an epidemic. Affordability of mobile internet is a burden to digital inclusion in disadvantaged communities, especially considering that the median cost of 1 GB of mobile data is 15.3% of monthly GDP per capita for the poorest income quintile [24]. We recommend that other initiatives seeking to implement DCT in low-income settings provide DCT free of charge to end users. Including ICT providers in the deployment of a DCT app is also crucial to ensure sufficient network coverage and, if applicable, remote installation of the app.

Other GPS-based DCT apps have been used in South Korea (Corona 100), China, Israel (Ha’Magens) or the USA (Private Kit) with varying success [9]. To date, few peer-reviewed papers on the effectiveness of these apps have been published. The impairment of GPS accuracy in urban settings, as seen in our results, is consistent with other studies [25, 26]. Although the Bluetooth-assisted GPS location-based DCT app improved accuracy in buildings, this came at the cost of a higher risk of recording false-positive hits when GPS was switched off. To reduce this risk, the interval between the timestamp of the last recorded GPS coordinate and the timestamp of coordinates based on Bluetooth pairing needs to be matched. High numbers of false-positive exposure notifications raise doubts about the usefulness of the app in the population, as reported for the Ha’Magens app in Israel [27], a location-based DCT app used as an automated tool for DCT.

Impairment of GPS accuracy in urban settings together with the other DCT limitations identified suggests using DCT to supplement traditional manual contact tracing. In case of an infection, the app users can voluntarily look at a map with their logged GPS coordinates, which can only be accessed together with a verified health official (e.g., contact tracers). The ‘recall problem’ has often been described as one of the weaknesses of manual contact tracing [28]. Listing the possible contacts during a period of time, e.g., the last 2 weeks, from memory alone is often insufficient. By using GPS coordinate logging of the last 2 weeks, prospective contact tracing is likely to be more accurate and effective. Another strategy that was reported as highly successful during the COVID-19 pandemic, but rarely implemented, is retrospective cluster-based tracing [29]. The intention is to recognize infection clusters early by asking where people have been infected and thereby informing other people from the cluster who might have been infected but who are still presymptomatic or asymptomatic. One reason retrospective cluster-based tracing was not commonly used is that the additional tracing effort might be beyond the capacity of public health officials during the surge of a pandemic. Aggregated GPS coordinate logs of infected persons are able to reveal clusters of infection early and thus enable containment measures to be implemented sooner.

High trust in authorities and health officials is an important factor in achieving high rates of app uptake, acceptance and adherence [27, 30, 31]. Misinformation and poor transparency surrounding data privacy have led to some DCT attempts being unsuccessful [27]. South Korea applied DCT requiring high levels of sensitive data access (geolocation, medical records, camera, financial transaction), yet due to transparent and accurate information communicated by the government, the population demonstrated trust in the app, which boosted the effectiveness of contact tracing. When surveyed, 86% of the population stated that the government had done ‘a good job’ dealing with the pandemic and that their country was more united as a result [32]. In contrast to many East Asian societies, there is little evidence that this level of privacy intrusion could be replicated within European societies due to their historical skepticism toward state surveillance [7, 9, 33, 34]. This highlights that social groups might differ in their judgment of fundamental considerations based on cultural differences, social preferences or individual risks. Deliberation of privacy concerns and DCT effectiveness (voluntary vs. mandatory, decentralized vs. centralized approaches) should be considered within the specific context of epidemics. Given epidemics with high case fatality rates, high numbers of infected individuals [35] and a collapsing health system (e.g., the Ebola epidemic of 2014–2016), it is likely that measures for effective disease containment could be valued higher than data privacy concerns by the population. Perceived personal threat and lack of personal control are positively related to the acceptance of surveillance technologies [36, 37]. This is consistent with our study results, where a high overall acceptance of a DCT app was identified in a cohort of health workers who had previously also been confronted with a humanitarian crisis during the Ebola epidemic in 2014–16. In spite of this, however, we still detected a discrepancy between high reported overall acceptance of DCT and low levels of adherence to app use during the study. The high rates of switched off GPS, seen in the results of this study, in large parts can be explained by increased battery consumption in combination with limited energy supply (some participants could not use their power generator due to shortage of fuel during the study). Resolving the burden of limited internet access would increase the number of hours the DCT app tracked by 11.3%. Additional factors such as smartphone quality and technical experience have been identified to explain reduced compliance.

The detected reasons for the finding that only 1.1% of the maximum Hours Recordable (mHR) during the study period was recorded by the participants using the DCT app (tHR) are consistent with the known drivers and barriers to uptake of DCT, which can be classified into four categories: motivation, access, skills and trust [38]. The main kinds of motivation are changes in the perception of risk and perception of effectiveness of DCT, convenience of app use (e.g., battery consumption, mobile data), the bandwagon effect and whether DCT is mandatory or voluntary, which have a huge influence on app uptake. The accessibility (affordability/availability) of the necessary equipment for DCT (internet, electricity, hardware and software) determines whether people can participate in DCT. A sufficient level of technical or literacy skill and an easy-to-understand format of information are required to understand what DCT is and how it works. Trust in entities associated with DCT, such as governments, corporations (Google, Apple, etc.), data architecture (centralized vs. decentralized) and data security biases the decision of people to use DCT.

The decision to participate in DCT or not depends on the complex interaction of these factors. A dependency on individual control of many of these factors has been reported [38]. Because during an epidemic the sense of loss of control is likely to increase, an essential question with respect to people’s decisions to participate in DCT is: do people believe that their choices and actions create a positive change? Large-scale surveys [39] found that individual control is correlated with perceived knowledge. Giving people information about (effective) DCT influences their sense of control and can lead to high levels of uptake and compliance because of the feeling of contributing to containing the epidemic.

To ensure effective communication at the community level and to reach vulnerable and disadvantaged groups, we implemented a feature in the DCT app where relevant information could be shared:

Managing misinformation is an important component of a response strategy because, as seen during the Ebola epidemic in West Africa in 2014–16 [1, 39] and the COVID-19 pandemic [40], misinformation can inhibit effective epidemic containment. To meet the functional criteria and minimize the bureaucratic burdens for the implementation of a DCT app, we identified the involvement of health authorities (e.g., MoH) as an important operational success factor. Management and maintenance of a DCT app requires a considerable amount of qualified personnel, skills and experience from local health officials and institutions. Additionally, proactive engagement of community leaders and religious or cultural figures should be encouraged for high rates of acceptance throughout the population [38].