After screening by title and abstract, out of the 1054 identified articles after eliminating duplicates, 134 were included for full-text reading, and 16 met the criteria for the extraction of relevant information. A manual search provided 4 additional studies; thus, 20 were finally eligible for inclusion (figure 1).
Fig. 1 Scoping Review Flow Diagram following the PRISMA 2020 statement proposed by Page et al. (2021)
Of those selected, we found: 10 narrative reviews (Grantz et al., 2020; Klingwort & Schnell, 2020; Mali & Pratap, 2020; Malik et al., 2020; Marabelli et al., 2021; Mbunge, 2020; Park et al., 2021; Roche, 2020; Röösli et al., 2021; Scott & Coiera, 2020; Shachar et al., 2020), 5 original articles (Casiraghi et al., 2020; Hisada et al., 2020; Marabelli et al., 2021; Moss & Metcalf, 2020; Ravizza et al., 2021), 2 case studies (Gulliver et al., 2020; Sáez et al., 2021), 2 systematic reviews (Mbunge et al., 2020; Wynants et al., 2020), and 1 rapid review (Anglemyer et al., 2020). Their main characteristics are shown in Table 3.
Table 3
Main characteristics of included studies
First author and year
|
Country
|
Goal
|
Design
|
Triage/
Patient risk prediction
|
Contact tracing
|
Anglemyer, Andrew (2020)
|
New Zealand
|
To assess the benefits and harms of digital solutions for identified positive cases of an infectious disease and to assess acceptability of this approach from qualitative studies.
|
Rapid Review
|
|
YES
|
Casiraghi, Elena (2020)
|
Italy
|
To develop an AI prediction model capable of processing clinical, radiological and laboratory data of patients related to COVID19 to predict their risk.
|
Original article: Development of a predictive ML-based computational computing system
|
YES
|
|
Grantz, Kyra H. (2020)
|
USA
|
To review the different applications for mobile phone data in guiding and evaluating COVID-19 response and its potential selection bias; to discuss best practices and potential pitfalls for integrating these data into public health decision making.
|
Narrative review
|
|
YES
|
Gulliver, Robyn (2020)
|
Australia
|
To present a six-stage model to evaluate and design best practice infrastructure to use big data in social policy. To provide informative and actionable technical guidance for social policy makers and researchers setting to use big data in their projects.
|
Case study
|
|
YES
|
Hisada, Shohei (2020)
|
Japan
|
To identify clusters of COVID-19 through web search query logs of multiple devices and user location information from location-aware mobile devices.
|
Original article: tracking the activity on some websites
|
|
YES
|
Klingwort, Jonas (2020)
|
Germany
|
To review weaknesses and limitations of tools such as smartphone contact apps to monitor the spread of COVID-19, to prove that no useful results can be obtained and suggest feasible alternative data sources for valid and population covering COVID-19 indicator systems.
|
Narrative review
|
|
YES
|
Mali, Suraj N. (2020)
|
India
|
To review opportunities and risks of AI applied to COVID-19.
|
Narrative review
|
YES
|
YES
|
Malik, Yashpal Singh (2020)
|
India
|
To overview prospective applications of AI model systems in healthcare settings during the COVID-19 pandemic.
|
Narrative review
|
YES
|
YES
|
Marabelli, Marco (2021)
|
USA
|
To study how IT has affected and will affect individual, organizational and societal practices during and after the COVID-19 pandemic. To develop a theoretical construct called "digital scars", defined as ethically problematic sociotechnical innovations that outlast emergency rollouts.
|
Original article: role of ubiquitous computing during COVID-19
|
|
YES
|
Mbunge, Elliot (2020)
|
Swaziland
|
To provide a comprehensive review of emerging technologies to address COVID-19 with emphasis on characteristics, challenges and country of domiciliation.
|
Systematic review
|
YES
|
YES
|
Mbunge, Elliot (2020)
|
Swaziland
|
To analyze potential opportunities and challenges of integrating emerging technologies for the implementation of COVID19 contact tracing.
|
Narrative review
|
|
YES
|
Moss, Emanuel (2020)
|
USA
|
To describe and consider the role of machine learning in social production of risk, the role of risk management in the effort to institutionalize ethics in the technology industry and its possible benefits during pandemic crisis.
|
Original article: risk management in Machine Learning
|
|
YES
|
Ravizza, Alice (2021)
|
Italy
|
To integrate the international framework of requirements to mitigate the known problems of mobile applications to monitoring and tracking and to suggest a method for clinical data collection that ensures researchers and public health institutions significant and reliable data.
|
Original paper: framework for epidemiological database creation
|
|
YES
|
Roche, Stéphane (2020)
|
Canada
|
To present an overview of the development of Contact Tracing and suggest a reflection on possible solutions for their ethical and sustainable deployment through a more active and transparent citizen engagement.
|
Narrative review
|
|
YES
|
Röösli, Eliane (2021)
|
USA
|
To review some of the risks of perpetuating biases using AI-based models to address COVID-19.
|
Narrative review
|
YES
|
|
Sáez, Carlos (2021)
|
Spain
|
To show the potential limitations that multisource variability may have for COVID-19 ML research on large international DRNs. To discover and classify severity subgroups using symptoms and comorbidities
|
Case study
|
YES
|
YES
|
Park, Sangchul (2020)
|
USA-South Korea
|
To present the main concerns over privacy involving tracing strategy through IT in South Korea.
|
Narrative review
|
|
YES
|
Scott, Ian A. (2020)
|
Australia
|
To describe several applications of AI relevant to COVID-19.
|
Narrative review
|
|
YES
|
Shachar, Carmel (2020)
|
USA
|
To expose legal and ethical concerns in AI applications to combat COVID-19 (privacy, human rights, equality and actors involved) and to give frameworks to guide stakeholders.
|
Narrative review
|
|
YES
|
Wynants, Laure (2020)
|
The Netherlands
|
To review and appraise the validity and usefulness of published and preprint reports of prediction models for diagnosing COVID-19 in patients with suspected infection, for prognosis of patients with COVID-19, and for detecting people in the general population at increased risk of COVID-19 infection or being admitted to hospital with the disease.
|
Systematic review and critical appraisal.
|
YES
|
|
Table 4 summarizes the different apps identified in this study. We use the term “bias” to refer to the systematic errors in a computer system causing an inclination or prejudice for or against a person or a group of people that can be considered to be unfair and cause a deviation from the expected prediction behavior of an AI tool, and “limitations” to refer to facts or situations that allow only some actions and make others impossible.
Table 4
DCTApps identified in the scoping review
APP
|
COUNTRY
|
TYPE
|
TECHNOLOGY
|
MANDATORY /OPTIONAL
|
REFERENCE
|
TraceTogether
|
Singapour
|
Contact Tracing
|
Bluetooth (BlueTrace protocol)
|
Optional
|
(Roche 2020)
|
The-Corona-Warn-App
|
Germany
|
Contact Tracing
|
Control smartphone
|
Optional
|
(Mbunge 2020)
|
COVIDSafe
|
Australia
|
Outbreak Response
|
Bluetooth (BlueTrace protocol)
|
Optional
|
(Gulliver et al. 2020)
|
Stopp Corona
|
Austria
|
Contact Tracing
|
Bluetooth
|
Optional
|
(Mbunge 2020)
|
BeAware App
|
Bahrain
|
Contact Tracing
|
Bluetooth and Global System for Mobile Communications technology
|
Optional
|
(Mbunge et al. 2020)
|
HaMagen
|
Israel
|
Contact Tracing
|
Bluetooth and GPS
|
Optional
|
(Mbunge et al. 2020)
|
bStayHomeSafe
|
China
|
Contact Tracing
|
Bluetooth, GPS and WiFi
|
Mandatory use
|
(Mbunge 2020)
|
CoronaApp
|
Colombia
|
Contact Tracing
|
Global Positioning System
|
Optional
|
(Mbunge et al. 2020)
|
Aarogya Setu
|
India
|
Contact Tracing
|
Global Positioning System
|
Mandatory use
|
(Mbunge 2020)
|
GH COVID-19
|
Ghana
|
Outbreak Response
|
Global Positioning System and GIS
|
Optional
|
(Mbunge et al. 2020)
|
CoronaMadrid
|
Spain
|
Symptom tracking
|
Global Positioning System
|
Optional
|
(Mbunge 2020)
|
Social Monitoring
|
Russia
|
Quarantine compliance
|
Global Positioning System
|
Optional
|
(Mbunge et al. 2020)
|
Yahoo! JAPAN App
|
Japan
|
Contact Tracing
|
WSSCI
|
Optional
|
(Hisada et al. 2020)
|
StopCovid
|
France
|
Contact Tracing
|
Bluetooth
|
Optional
|
(Roche 2020)
|
STOPV
|
France
|
Contact Tracing
|
Global Positioning System, Semantic Data, Epidemiological Data and Test Results
|
Optional
|
(Roche 2020)
|
Private Kit: Safe Paths
|
USA
|
Contact Tracing
|
Global Positioning System
|
Optional
|
(Roche 2020)
|
Covid Alert
|
Canada
|
Contact Tracing
|
Bluetooth
|
Optional
|
(Roche 2020)
|
a. AI systems developed for triage and PRP.
a. 1) Bias
One of the most relevant aspects addressed in the literature is data-related biases. According to a systematic review (SR) of COVID-19 prognostic and risk prediction methods (Wynants et al. 2020), there is a high risk of bias in the studies included due to a poor description of the population, which raises concerns about the reliability of their predictions when applied to clinical practice. An immediate exchange of well-documented individual participant data from COVID-19 studies is needed to develop more rigorous prediction models and validate the existing ones through collaborative efforts.
Our results identified different types of data-related bias:
a.1.1) Data source variability contributes to bias in distributed research networks of COVID-19 data sharing (Sáez et al., 2021), and they play an important role in data quality. The case study reported by Sáez et al. (2021) shows the limitations that multisource variability may have for COVID-19 machine learning (ML) research on international distributed research networks. They used the nCov2019 dataset, including patient-level data from several countries, to discover and classify severity subgroups dividing them into six types: 1) mild disease with no comorbidity; 2) elderly + severe pulmonary disease + comorbidity; 3) middle-aged + severe pulmonary disease + no comorbidity; 4) elderly + mild disease + no comorbidity; 5) elderly + severe systemic disease + comorbidity; 6) elderly + severe pulmonary disease + heart failure. The problem appears when this division is conditioned by data's country of origin. Groups 1 and 4 data were collected in China and Groups 2, 3, 5, and 6 in the Philippines. In the last case, data came from the COVID-19 tracker, owned by the Department of Health of the Republic of Philippines; in the case of China, data came mostly from patient reports. Due to these variations in the sources, results show some inconsistencies, limiting the model. Potential biases of multisource variability for ML can be generalized in large cross-border distributed research networks. How can we prevent such biases? Sáez et al. (2021) propose: 1) a routine assessment of the variability among data sources in ML and statistical methodologies could potentially reduce biases or extra costs 2) a complete data quality; 3) reporting data quality and its impacts as a routine practice in publications; 4) building consciousness about data quality and variability.
a.1.2) Casiraghi et al. (2020) developed an explainable PRP model for COVID-19 risk assessment aimed to avoid data bias. Their model was designed to be used in emergency departments for an early assessment of PRP in COVID-19 patients, integrating clinical, laboratory, and radiological data. The study carried out a comparative evaluation of different imputation techniques to manage the problem of missing data in the prediction for COVID-19 patients. However, the lack of a shared dataset hindered an objective comparative evaluation with the best models (Casiraghi et al. 2020).
a.1.3) There are biases in COVID-19 prediction models due to unrepresentative data samples, high probability of model overfitting, imprecise information on the study populations, and the use of a model that wasn't well suited for the task (Röösli et al., 2021). Models developed in elite and affluent academic health systems that are not representative of the general population lack external validity (Röösli et al., 2021).
a.1.4) The quick development of AI systems carries great risk due to skewed training data, lack of reproducibility, and lack of a regulated COVID-19 data resource (Röösli et al., 2021). Without comprehensive bias mitigation strategies, this can exacerbate existing health disparities. "The source code of any AI model should be shared publicly to ensure that the models can be widely applied, generalized, and transparently compared” (Röösli et al., 2021: 191).
a. 2) Limitations
A sufficient amount of high-quality data is crucial for the successful implementation of AI in COVID‐19 management. Designing practical AI‐based algorithms is challenging because of the huge and complex data that emerge as a consequence of the varied manifestations of the COVID-19 infection, ranging from asymptomatic to severe clinical disease (Malik et al., 2020). Moreover, the principal obstacle to implement these systems in the clinical context is the regulation of the data exchange obtained by the AI application. Additionally, AI-based algorithms can offer a binary answer to a specific question about the disease in context, but cannot offer alternative predictions (Malik et al., 2020). Finally, it is necessary to consider how meaningful and in-depth data can be generated at every point of healthcare activity (Malik et al., 2020).
a.3) Other ethical issues
Transparency in AI algorithms is essential to understand predictions and target populations, unrecognized biases, class imbalance problems, and their capacity to generalize emerging technologies across hospital settings and populations (Röösli et al., 2021). To ensure that models can be broadly applied, generalized, and compared, the source code of an AI system should be shared publicly, and regulatory frameworks should be created to facilitate data sharing (Röösli et al., 2021).
b. AI systems developed for DCT
b. 1) Bias
b.1.1) Uncontrolled application development could generate inadequate data collection and biases due to the loss of some data or an insufficient frequency of monitoring, which can lead to inability to compare data collected from different regions (Ravizza et al., 2021). Although it does not affect the core functionality of the app, it can influence further use of the collected data: most ML models have relied on Chinese data, which can limit scalability to other populations (Scott & Coiera, 2020).
b.1.2) The media alter the nature of searches, producing biases in areas of potential clusters. Whenever the media reports a location of a positive COVID-19 patient, many people who are close to the informed location ask for additional information related to COVID-19 (Hisada et al., 2020).
b.1.3) DCT Applications (DCTApps) pose a high risk of discrimination, especially to affected people (Mbunge, 2020). Specifically, Internet-of-Things (IoT) based DCTApps collect data from the entire population in real-time which is later analyzed to map COVID-19 hotspots. Such data include ethnic information, demographic details, and socioeconomic status, which can influence the allocation and distribution of COVID-19 resources potentially leading to discrimination.
b.1.4) False negatives are an obstacle and may be deliberately generated because infected people do not want to reveal their true status (Klingwort & Schnell, 2020). To overcome this problem, the detection of relevant contacts should be refined as the issue is fundamentally a problem of microscale spatial analysis. Applications must develop the microgeographic analytical capability to specify what kind of proximity constitutes a sufficient contagion risk to trigger a notification (Roche, 2020).
b.2) Limitations and technical problems
b.2.1 ) Accuracy
The most widely proposed type of COVID-19 application uses Bluetooth signals to track encounters with people diagnosed as infected after the encounter; the accuracy of automatic DCTApps suffers from Bluetooth-based measurement errors (Klingwort & Schnell, 2020). These errors are due to the devices’ different signal strengths and the fact that signal is not transmitted in all directions. Characteristics of the physical environment (windows, walls, or doors) can affect the range of discoverable devices. In addition to the four efficiency conditions (mass adoption, well-equipped population, numerous diagnostic tests, and fair and transparent uses), these monitoring applications have many reliability limitations, especially in the Bluetooth reading forecast and the calibration (Roche, 2020). This can add noise and produce many false positives.
b.2.2) Data-related problems.
Most applications have not reached operational maturity (Scott & Coiera, 2020) and their effectiveness has not been proved (Anglemyer et al, 2020). Even modelling studies provide low-certainty evidence of a reduction in secondary cases if CT is used together with other public health measures such as self-isolation. Cohort studies provide very low-certainty evidence that digital DCT may produce more reliable counts of contacts and reduce time to complete DCT (Anglemyer, 2020). The performance of emerging technologies is not yet stable in account of the lack of availability of a sufficient COVID-19 dataset, the inconsistency of some of the available datasets, the non-aggregation of the dataset, and missing data and noise (Mbunge et al., 2020).
DCTApps may use the IoT to transfer data to national health systems. However, they are not globally standardized, and they face a lot of problems based on interoperability (heterogeneity of connection standards and communication protocols, data semantics, formats, different operating systems, and programming languages). Consequently, each country has developed its own app. Data formats and structures should be standardized to avoid noise, prevent incomplete data, and improve its quality (Mbunge, 2020; Mbunge et al., 2020). Determining a standardized list of data, symptoms, clinical signs, risk factors, and comorbidities associated with coronavirus can contribute to ensure compatibility of databases between regions and countries and to improve interoperability (Ravizza et al., 2021).
DCT becomes less effective when dealing with asymptomatic individuals since symptom checkers and apps rely on pulse, temperature, and sleeping patterns (Hellewell, 2020, cited in Mbunge, 2020). Due to built-in privacy mechanisms, the resulting data for scientific research based on these applications is limited to counts of positive or negative encounters from selective populations, where the odds of encounters cannot be calculated (Klingwort & Schnell, 2020).
b.3) Other ethical issues
b.3.1) Privacy concerns
The use of DCTApps raises ethical, legal, security, and privacy concerns (Roche 2020). To be acceptable, this interference with fundamental rights must be justified, reasonable, proportionate, and politically consensual. DCTApps provide little or no privacy to infected people and require them to disclose their data, raising difficult issues of consent, privacy, ethics, and trade-offs between public and private goods (Scott and Coiera, 2020).
DCTApps violate the security, confidentiality, integrity, data availability of COVID-19 patients and contact persons, which can sometimes cause mental health issues like stress, anxiety, or depression (Mbunge, 2020). Apps like TraceTogether, COVIDSafe, or BeAware support access to multiple data access points and the monitoring and surveillance of infected or isolated people, which threatens the security of public health data, and may imply a violation of privacy (Mbunge 2020).
The study of Park et al. (2020) in South Korea recreates privacy-related problems (figure 2).
Fig. 2 Authors summary of a case example about privacy-related problems described by Park et al. ( 2020).
Instead of disclosing data to the public, information could be used to sanitize establishments, potentially avoiding stigma and business decline. That is, instead of publicly revealing the precise locations of an infected individual, less granular data could be revealed, with the same effect on tracking and quarantine (Park et al., 2020).
The correlation of data, the exchange of information, and the ability to extract information from different entry points contribute to the increasing fragility of the anonymization of data. This anonymization is even more fragile when information is collected over time and through data cross-referencing (Roche, 2020). The deactivation of DCTApps must be programmed so that monitoring does not continue beyond the health emergency and is not tacitly established as standard practice. Otherwise, risks of mass surveillance could arise.
b.3.2) Lack of regulation
There are no specific regulations for DCTApps. However, their use of data, access, or privacy has been adapted to international, national, and state laws such as the European Union’s General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA). These legal frameworks can be adapted to help address concerns about privacy, human rights, due process, and equality (Shachar et al., 2020). In certain countries like the US, the lack of state regulation makes it more difficult to guarantee that these applications follow ethical standards (Shachar et al., 2020), and there are no global WHO guidelines on health data shared and transmitted via 5G technology (Mbunge et al., 2020). Though some countries’ regulations protect citizens better, potential “digital scars” are left in society as long as the governments and private institutions continue having long-term and unlimited access to citizens’ data for surveillance purposes.
b.3.3) Consent
The efficacy of DCTApp depends on the level of population uptake, its ability to accurately detect infectious contacts, and the extent of adherence to self-isolation by notified contacts (Scott and Coiera, 2020). DCT must be handled with care: these technological solutions are proposed as the only tool available to ensure a process of deconfinement, a requirement that would make it a sine qua non condition accessible to police control. The risks of seeing such an established form of socio-spatial “triage” and patients and certain categories of the ostracized population are huge (Roche, 2020).
Although DCTApps use WiFi, GPS, or Bluetooth protocols to monitor people’s movement, users have the right to opt-out and configure their devices, jeopardizing the monitorization of positive cases (Mbunge, 2020). DCTApp should allow people to practice withdrawal of consent (Mbunge, 2020), as problematic uses of technologies may well remain once the pandemic is over. This can potentially advantage powerful groups that can obtain financial and political benefits from perpetuating the use of IT while having questionable effects on society (Marabelli et al., 2021).
Privacy issues related to forcing a population to use an app can lead to much lower coverage rates (Klingwort & Schnell, 2020). However, we find opposite scenarios in countries that have not developed any specific app. Brazil, for example, has increased its technological surveillance in order to minimize the COVID-19 transmission chain (Mbunge et al., 2020). This enforcement of massive surveillance can raise issues about power, abuse, and data exploitation.
c. Health disparities and social determinants of health in AI systems developed for triage and DCT for COVID-19
c.1) Racial disparities
Health disparities are related to the emergence of biases in ML systems in the US context where Black and Latinx communities have been the most severely affected by COVID-19 (Moss & Metcalf, 2020). This is due to long-standing disparities in health outcomes for these communities, the impact of environmental determinants of health, and the disproportionate number of workers whose jobs do not allow them to stay at home (Moss & Metcalf, 2020).
c. 2) Biased data
The reliance on AI may create a false sense of objectivity and fairness (Röösli et al., 2021). The pervasiveness of biases is a failure to develop mitigation strategies and has exacerbated the risk of existing health disparities, hindering the adoption of other tools that could actually improve patient outcomes. As an example, the Medical Information Mart for Intensive Care (MIMIC) is a publicly available, de-identified, and broadly studied dataset for critical care patients. A MIMIC-equivalent for COVID-19 from diverse data sources could incentivize urgently needed data sharing and interoperability to enable diverse, population-based tailored therapy—a step that could decisively reduce biases and disparities in healthcare while bolstering clinical judgment and decision-making. One of the main methodological problems is the selection process (Klingwort & Schnell, 2020). The sample of the population using the application will not be random, and subpopulations with a higher prevalence of undetected infections will likely have lower coverage. In addition, models that include comorbidities associated with worse outcomes in COVID-19 may perpetuate structural biases that have led to historically disadvantaged groups disproportionately suffering those comorbidities. To avoid further harm to minority groups already most affected by COVID-19, resource allocation models must go beyond a utilitarian foundation and must be able to identify needs amongst these patients (Moss & Metcalf, 2020).
c.3) Socio-economic disparities
In DCT, the ability to make use of notifications to minimize one’s own risk by self-quarantining is far too dependent on one’s personal wealth and capacity to afford to stay at home (Moss & Metcalf, 2020). DCTApps’ designers must be attuned to the context of social life in which such systems can produce harmful, difficult-to-foresee effects that replicate or amplify pre-existing inequalities. Attending the contextual use of such a system could collectivize risk by identifying and emphasizing the necessary forms of social support for self-quarantine and medical care: adequate sick leave and quarantine leave policies, robust testing, and the economic relief that targets individual workers over large companies.
During the pandemic, ML has been involved in the production and distribution of risk through society (Moss & Metcalf, 2020), generating risks and its uneven distribution in society. Many of the predictive surveillance algorithms used in DCT control focus attention on populations where bias is very present, especially in highly racialized or lower-income populations (Moss & Metcalf, 2020). In this sense, ML can be epidemiologically effective, while unethical.
c. 4) Unequal accessibility
AI-based global health initiative is recommended, since AI-based approaches may not be accessible in countries with limited resources (Malik et al., 2020). Regarding socioeconomic disparities and the digital gap, the lack of population coverage can leave certain populations at risk (Ravizza et al., 2021). Digital solutions can exacerbate existing disparities between those who do not have access to smartphones or who live in areas without connectivity, because of ethnicity, socio-economic status or age (Anglemyer et al., 2020), with equity implications for at-risk populations with poor access to the Internet and digital technology. Digital deserts or data poverty in certain geographical areas are concerning, especially because the effectiveness of DCTApps depends on their massive voluntary adoption and a systematic screening (Roche, 2020). Across country borders, the health gap and inequalities in healthcare pose a problem for the integration of emerging technologies. Even in developed countries, risk groups may not have access to broadband, smartphones, or wearable technology. For a community to benefit from this technology, most people need to be equipped with mobile devices. This applies to only 80% of the US population, 65% of Russians, and 45% of Brazilians (Marabelli et al., 2021). Children, elderly, or individuals with fewer resources are excluded from the stored information (Grantz et al., 2020).
c.5) Workforce and Information and Communication Technologies infrastructure
Developing an app and maintaining the system requires a specific workforce and a consistent Information and Communication Technologies (ICT) infrastructure that may be lacking. Some countries may struggle with the technological infrastructure, especially in countries with a high incidence (Chad or the Central African Republic) where ICT infrastructures are very poor. These factors can hinder the development of technological innovation policies as part of their response to COVID-19 (Mbubnge, 2020).