Data governance and ethics in digital health surveillance for emerging infectious diseases

doi:10.21203/rs.3.rs-3993737/v1

Download PDF

Research Article

Data governance and ethics in digital health surveillance for emerging infectious diseases

https://doi.org/10.21203/rs.3.rs-3993737/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Epidemic intelligence, and in particular, its component of digital health surveillance, combines multiple large, heterogeneous datasets, often by using artificial intelligence (AI) systems to detect, monitor, and assess threats relevant to public and animal health. This could raise significant ethical issues regarding data sources, natural language processing, user privacy and consent, among others. The European Commission is highly engaged in how European projects using AI for health data and digital health surveillance comply with the General Data Protection Regulation and ethical principles.

This work aimed to better understand the governance of data in the H2020 MOOD (Monitoring Outbreak for Disease Surveillance in Data Science Context) project. The authors also studied the perceptions and views of researchers on ethical risks and suggested actions to mitigate these risks in an international multisource Big Data Analytics and One Health project.

First, a data mapping approach was used to determine the origin and destination of the data in the project. Participatory observations were conducted to understand the data scientists at work. Information was also collected through a qualitative study using semi-structured interviews with eight project researchers ranging from data scientists to epidemiologists and ethics experts; a quantitative survey of all consortium members complemented this process.

Big data and AI systems have enormous potential for strengthening healthcare delivery, including deploying different public health interventions such as disease surveillance, outbreak response and health system management. However, some risks and constraints could hamper the reliability of data analysis and AI systems, such as the deidentification, lack of privacy, compliance with Twitter Application Programming Interfaces terms of use, and the risk of reproducing bias and stigmatisation of minorities. Our findings suggest that few researchers could be reluctant to work and establish action to mitigate ethical risk depending on the approach used in ethical counselling for European and transdisciplinary projects. The philosophical and comprehensive approach to ethics is judged softer when comparing the legal and more constraining requirements to comply with the law.

Using Big, multisource EI data in a One Health framework requires consideration of strong ethical principles that safeguard users’ privacy and constant ethical support for researchers.

Ethics

digital health surveillance

epidemic intelligence

algorithm bias

artificial intelligence

One Health

Developing a common infrastructure for disease risk analytics that stores large datasets is a major challenge in terms of governance and compliance with the legal framework in European countries
Managing multiple internet-based sources for event-based surveillance highlighted important ethical issues such as consent, privacy and risks of biases that may lead to unintended consequences in the health surveillance field
Researchers should comply with Ethical requirements based on the European, ethical and General Data Protection Regulation (GDPR) which has been transformed in deliverables into a dedicated Work package.
Mapping the data flow used to construct a future common prototype platform tailored for monitoring outbreaks and disease surveillance in a One Health framework allows for understanding governance issues and challenges that may arise when assembling big data
Project researchers' and ethics experts’ perceptions and views of these risks, to their experiences grounded in everyday practices of AI-related tasks, provide a better understanding of a future framework for assessing ethics risks in digital health surveillance projects.

The health-related content published on the Internet, through online news, ministerial and health agency websites, scientific literature and social media such as Twitter (now X), can be a relevant source of information on new and emerging infectious disease outbreaks (1–4). This heterogeneous and multisource Big Data is especially valuable for those working in the One Health framework and conducting epidemic intelligence (EI) activities in health (PH/AH) agencies; the EI is defined as all activities related to early identification, verification, analysis, assessment, and investigation of health threats (5). This multisource Big Data are useful for the early detection and monitoring of disease outbreaks, public health surveillance of mass gatherings, and investigation of public attitudes and behaviours related to disease control measures as a core of EI activities (2, 6–9). For example, by searching, filtering and visualising events of public health interest from unofficial sources, event-based surveillance (EBS) became a relevant part of EI services in Europe (10). Thus, EBS can be understood as “the inclusion of numeric and Big data, particularly from social and online media channels, news aggregators or other informal internet-based sources” to detect and notify threats of infectious disease emergence with epidemic potential (4, 5). Each dataset used in the EBS has its characteristics and proper value. Although raw, they are often qualified as unofficial, unstructured and not verified. The EBS is considered a complementary component to traditional, indicator-based surveillance (IBS) which relies on the collection and analysis of data on disease outbreaks and other health-related indicators using formal data sources, such as public health and animal health agencies (PHs/AHs), usually available in a structured format (5).

In EI, particularly for EBS, artificial intelligence (AI) is used in information extraction modelling through deep and machine learning algorithms that provide real-time analysis; trends in socioeconomic patterns or climate; and disease nowcasting and forecasting, which can aid in risk assessment of identified potential threats. Final reports with potential threats identified from EBS are also often generated by automated machine-based processes and, in certain cases, by human analysts or subject matter experts (11).

Big data analytics, particularly in health and medicine, embeds several risks, such as information overload to store and (real-time) process the data, data quality, heterogeneous data, data privacy and security (12). Several guidelines and recommendations provide generic ethical principles. For public health surveillance, the World Health Organisation underlined the importance of considering the balance between public health and the benefits of surveillance and ethical and legal concerns about individuals and communities concerning privacy and autonomy. Although AI systems intervene in the social world as efficiently as possible, the civil liability of the AI regime could be clearer. For example, it is unclear who is responsible for ethical, social and economic issues, controls the risk associated with the AI system, or what code, inputs or data ultimately cause a “damaging operation”. The European Parliament recommends clarifying that the research process does not harm communities or individuals, surveillance results or automated decision-making (13). In clinical trial research, the European Commission's (EC) General Data Protection Regulation (GDPR) guarantees the framework of individual consent. However, in research using social media posts for public health purposes, reusing the terms and conditions of the Twitter Application Programming Interface (API) may not be sufficient to safeguard consent and privacy (14).

For example, the use of AI in automated EBS data processing does not guarantee that data cannot be lost or prone to outages in the onlinesources (e.g., social media). Scholars have noted that we need to understand the properties and limitations of the datasets we use for analysis, even if they include millions of data points, i.e., big data. Hence, this does not mean that big data is random and representative, as AI systems proceed by reduction and could lead to the distortion of facts (15). Big data surveillance may have low levels of transparency and comprehensiveness, and the infrastructure used may be structurally asymmetrical (16).

The World Health Organisation (WHO) has therefore established core principles to promote the ethical use of AI in health: i) protect autonomy; ii) promote human well-being, human safety and the public interest; iii) ensure transparency, explainability and intelligibility; iv) foster responsibility and accountability; v) ensure inclusiveness and equity; and vi) promote AI that is responsive and sustainable. The European Commission may check compliance with the legal aspect and respect for the ethical framework and guidelines to foster a trustworthy AI process and mitigate unintended consequences in the health system (17).

MOnitoring Outbreaks for Disease surveillance in a data science context (MOOD) (https://mood-h2020.eu/) is a multidisciplinary, international project aimed at developing tools and services for European public and animal health (PH/AH) agencies and EI services to foster early detection, assessment and monitoring of emerging infectious diseases. The project involves several partners from the EU and non-EU countries, thus addressing the challenge of cross-sectoral data sharing for EI purposes and a One Health approach framework involving human, animal and environmental health data. The main objective of the project is to construct a data-driven translational research MOOD prototype platform; that is freely available or at a low cost for use by EI practitioners at PH/AH.

The project coincided with the beginning of the COVID-19 pandemic in January 2020, when there was a high demand to generate knowledge and a better understanding of the transmission of SARS-CoV-2, its control and exit strategies. In the MOOD project, several steps and decisions guarantee that researchers are compliant with the requirements and follow the ethical guidelines of the European Commission (17) to safeguard and protect the data related to persons (all people or groups of people who have personal information stored in the platform or for whom personal information can be easily deducted). The steps followed by the project are as follows:

1) Description of the project’s dataflow;

Designation of a Data Protection Officer (DPO) by each institution - a member of the project;

Data management plan and registry of assessed risks with mitigation measures, for each task, updated by each institution and task leader - members of the project;

Description of the data processing and curation by each institution - a member of the project;

When collecting data, each researcher -a member of the project should assess how others may exercise access rights to their data, for example, rights for modification or deletion;

Twelve deliverables were established in line with the requested ethics and data protection requirements for the project.

Although ethical principles that apply to the use of AI can provide safeguard and normative values in health science, little is known about data governance and ethical issues at stake in programs related to multi-source big data analytics for EI. Hence, our objective is to present the ethical issues that may arise in European programs supporting big data analytics through data science and AI in the domain of emerging infectious disease surveillance, in particular; 1) the flow of data 2) their governance and 3) perceptions of ethical issues from scientists. MOOD partners are developing a metadata catalogue to list all the data sources and their respective data protection measures. This catalogue is available on the MOOD website at https://mood-h2020.eu/.

We use the MOOD project as a case study, and we show that building a single prototype platform using Big Data for the surveillance of outbreaks with multiple infrastructures requires overcoming several challenges in the governance of data and AI systems such as the liability and interoperability of those systems. Putting together different sources of data for risk maps and disease analytics requires sharing knowledge among researchers who confront ethical issues related to the risk of reconducting bias, reproducing stereotypes or the question of privacy while the quick shift of the internet ecosystem and its rules need adaptation and innovative ways to do research and surveillance. To construct a sociotechnical system using algorithms for modelling disease, for example, researchers need to shift AI ecosystems toward equitable and accountable AI. Kondylakis et al. addressed interesting insights into complying with the GDPR and ethical guidelines when creating an AI cloud-based repository for cancer care provision, an initiative involving five EU projects focused on health imaging (18). To the best of our knowledge, our study is the first to provide researchers with insights and evidence about the governance of data and ethics knowledge in a One Health framework.

2.1 Overview of the MOOD project

The workflow of the MOOD project is divided into eight work packages (WPs), each with different roles and tasks (Supplementary file 1). WP1 provides the interface with stakeholders and potential users of the project tools and services, including understanding their needs, maintaining communication and ensuring the usability of the project outputs for different international, European and national PH/AH agencies. WP2 identifies disease indicators, proxies, and data to feed the MOOD prototype platform for major airborne, vector-borne, food and waterborne infectious diseases, including an unknown disease (disease X) and antimicrobial-resistant bacteria of interest to Europe. WP3 secures the data pipelines, merging, standardisation and uniform multisource data on outbreaks, climate and environmental changes, vector and host distributions, and antimicrobial resistance data. WP4 develops disease-specific and generic methods for the analysis and modelling of outbreak data concerning climate, environment and global change indicators and adapts the models based on user needs. WP5 develops the EI MOOD prototype platform as a sustainable open-source tool based on user needs. The supporting WPs (WP6-7) ensure the dissemination of information and impact assessment and capacity-building with potential end-users in the EU and beyond. Finally, WP8 identifies, reports and solves ethical issues encountered by researchers in their daily use of surveillance data, algorithms and AI and coordinates the ethical deliverables according to the requirements of the EC. The Ethics and Data Protection (DP) Board, which includes members composed of the researchers of the project and two consultants, experts in ethics and DP, also ensures that the activities concerned fulfil the data management plan and the ethics and DP plan, identify additional ethical risks and propose solutions.

2.2 Qualitative analysis

The qualitative analysis was conducted using stakeholder mapping and semi-structured interviews, as described below.

2.2.1 Mapping

The MOOD project uses multiple datasets ranging from environmental to epidemiological data. The aim of mapping the MOOD program data flow was to show the relational view of the data from their origin to their convergence in the MOOD prototype platform. The mapping allows one to have a better understanding of where the data originated from and where they go.

Design. The ethical risks related to data manipulation in digital epidemiological surveillance can appear at different levels of the work done within the MOOD project. To better visualise these data, we mapped the different elements that make up the MOOD database and their connections.

Recruitment and sampling. The parameters that have been considered are the data sources (institutions), the inputs, the type of data, to which they provide it, the type of work done with it by the WPs, the outputs and finally, the ethical risks identified by research scholars and data scientists working on the MOOD prototype platform.

Data collection. We analysed interviews, informal discussions, internal MOOD documents and personal notes. Feedback from project partners also allowed us to define the level of detail of the mapping and to check the validity of certain information.

2.2.2 Semi-structured interviews

Design. We follow and participate in conferences, meetings, and conversations about the data and tools that will be used to integrate the MOOD prototype platform. We performed participant observation in a data science laboratory to better visualise and understand the social practices involved in coding. The participant observation allows us to recruit people according to their level of interaction with the data analytics and ethical compliance. We established a semi-structured interview guide with limited questions.

Recruitment and sampling. Eight semi-directed interviews were conducted between mid-March and early May 2022. Participants were invited to answer questions during face-to-face meetings (with OT and LM) or remotely, according to where they worked in Europe. Consortium members are asked to provide informed consent as their participation was foreseen in the Grant Agreement. The key informants were involved in data science activity, leaders of tasks (risk mapping, visualisation, prototype platform design, and ethical adviser) or experts on ethics. These are members of the consortium or independent advisers hired for the programme.

Data collection. The Interviews were performed in English or French. We recorded the audio of each interview, and notes were taken. The remote interviews were performed via Microsoft Teams. The average duration of the interviews was 50 minutes. The participants were asked about the following main themes (Supplementary File 2) following a detailed interview guide (Supplementary File 3): 1) data collection, 2) data privacy, 3) data accessibility, 4) Artificial Intelligence and big data, 5) ethics standards, 6) risks of stigmatisation and discrimination, 7) Bias and inclusivity, and 8) Common ethical issues.

Data analysis. The interviews were transcribed and imported into NVIVO12© (QSR International version 12). We followed a coding method that is used in grounded theory analysis and is designed to identify emerging themes. This codebook helped to construct the survey by pre-identifying the priority themes (Supplementary file 4). All interviews were pseudonymised. The most illustrative narratives will be used further to illustrate the researcher's experience in data governance and AI health surveillance systems.

2.3 Quantitative analysis

The quantitative analysis was performed through a survey.

2.3.1 Survey

Design. The results of the interviews allowed us to develop a survey with key headings. The purpose of this form was to assess researchers' knowledge and issues related to the ethics of big data and AI in epidemic surveillance.

Recruitment and sampling. We invited MOOD partners through an intern mailing list to complete the survey form. The mailing list included 132 individuals. The people included in the interviews were partners involved in different WPs of the project or researchers manipulating the data within MOOD, ensuring the position of the leader of the data management task or who was responsible for the ethical aspect of the data.

Data collection. The survey was administered with Google Forms, which has a convenient interface and multiple possibilities. An explanatory introduction, the definitions of certain terms and a commentary box at the end were added to help understand the survey. The data collection was anonymous, and no email was collected. The questions varied in type and requirement. They included single-choice, multiple-choice, linear scale, and single-choice grid options. Some questions were optional for participants who did not have a relevant section (N/A). The survey completion duration was approximately ten minutes for 41 questions divided into seven sections. The seven sections were: 1) General Information, 2) Artificial Intelligence, 3) Data, 4) Open Science, 5) Ethics and Risks, 6) Ethics Board, and 7) Code of Conduct.

The online survey was released by email via the "mood-partners" mailing list which included the one hundred and thirty-two partners involved in the MOOD project (June 2022). No personal data were collected, and the answers were automatically anonymized via Google Forms. Three reminders were sent approximately weekly and a QR code directed to the questionnaire was also signed during a meeting of MOOD partners.

Data analysis. The responses collected by Google Forms were automatically transferred to a Google Sheets spreadsheet, allowing the data to be analysed quantitatively. Therefore we were able to analyse the answers to the questionnaires and construct graphs to better visualise the data.

3.1 Qualitative

3.1.1 Mapping

Mapping the MOOD data flow allowed one to have a better understanding of the actor-network, i.e., the data, institutions, tools and potential ethical risks (Fig. 1). Reading from top to bottom, several institutions such as the European Food Safety Authority (EFSA), the World Organization for Animal Health-World Animal Health Information System (WOAH-WAHIS) and the European Centre for Disease Control and Prevention (ECDC), handle surveillance platforms where they provide data (upon request) from databases such as the European Surveillance System (ECDC-TESSy). These sources provide aggregated epidemiological data on disease cases and outbreaks, most commonly per NUTS level 3 (Nomenclature of territorial units for statistics). Several datasets from the pipeline are distributed and converged into different activities and tasks that involve a high level of collaboration between data providers, agencies, partners and researchers and among researchers themselves. The incoming data are then handled by different WPs (here 2, 3 and 4 because they are the main data handlers); then they perform different types of work on them such as models and risk maps, etc. The WP5 deals with data storage, so we decided to differentiate it from the others because it manages all the services provided through a MOOD prototype platform. We identified several main activities in the actor-network: disease profiling and mapping, modelling and machine learning, and computer coding and programming. The actor-network analysis also showed that different actors in charge of several activities are intertwined with different data and databases, mainly covariates, surveillance (case/outbreak) data, genomic data, and media and social media data (Fig. 1). As Fig. 1 shows, the data flows represent the institutions providing data to the MOOD prototype platform during the study.

We identified several potential ethical risks at this level of work (Fig. 1), such as the risks of misuse when data moves from one place to another, and the data flow explicitly asked us to check compliance with the FAIR principles. Other risks, such as deidentification, exist when managing personal data (identity records or GPS data) despite an aggregation process, specifically when only one person in the area matches the targeted profile. Finally, this work produces different types of outputs linked together to feeds the MOOD prototype platform and poses issues with the legal framework when the data moves from one node to another.

3.1.2 Semi-structured interviews

Data sources and formatting issues

Epidemiologists and data scientists use several public and animal health databases of European agencies and international organisations related to prevalence data on zoonotic diseases, animal vectors, hosts and pathogens, animal health situations worldwide, data and information on microbiological agents in foodstuffs, country data sets on foodborne outbreaks and antimicrobial data, covariate data related to the environment and climate. EBS data are retrieved from mainstream newspapers, journals, social media, news aggregators, and automatic tools and platforms such as ProMED or PADI-Web and public repositories.

The data can have multiple formats, as stated by one respondent: "We have data that are not pre-processed at all; in fact, they are raw data" (AT95).

It reveals that some data arrives raw, i.e. non-anonymous, to researchers. A project partner further suggests that this is a problem, saying:

"There are two ways or the source... The organization that holds the epidemiological data gives the raw data, and there should be a problem, and that is when it is up to the modeller to anonymize. In addition, I imagine there could be a regulatory problem here because normally the organisation that holds the data should, in a responsible way, anonymise" (YN76).

Data processing and associated risks

An interviewee also admitted that the risks of reidentification, in particular by cross-referencing, exist when he announced:

"I think that if I searched, I could find the association or the owner of the poultry farmer" (AT95).

The anonymization of unofficial data collected for the program used different scales and followed the rule of coding generated by the community of practice:

"We try to aggregate by country, possibly by region"[The goal is] "not to make the tweets public, but to make public the tweet ID or just a number to find the tweet" (TC62).

However, the tweet ID is personal data that can make people identifiable, as one ethics expert highlighted during one special meeting.

An algorithm was used to anonymise the Twitter data, but a margin of error can exist, patterns that can identify users may remain:

"We have no guarantee that it works 100% [...] We can leave the names of people who are identified" (TC62).

The following risks of deidentification of the data are also underlined by ethics experts:

"There is a risk of merging the data that he [the researcher] collects with already existing data and therefore [...] if the databases or models that they have created are used in a broader context, it can lead to problems they do not suspect" (NM44).

Machine learning

Machine learning for automated information extraction often derives from a black box approach that lacks clarity about causal connections even for those who are familiar with the process; a researcher stated:

“On the other hand, it [the algorithms] learns, it builds a model. The model is mathematical, and based on deep learning; it is not too much; it is a bit of a black box. We do not truly know how it learns because it does it completely automatically" (TC62).

Another one admitted that:

"If you cannot explain what's going on in the black box, you should not use it; that is the basis" (HE47).

Data leave a trace

One of our respondents recalled that "one thing that comes with digitalisation is that everything leaves a trace […] it is impossible to erase data today" (ML50) about the right to withdraw data from individuals. Regarding unofficial data collection, a researcher says, "If a user removes their tweet, well, it stays in our database actually” (TC62).

Lack of inclusiveness

These interviews allowed us to identify certain biases such as the problems of non-representativeness of unofficial data and keywords chosen in a few Indo-European languages: "people who express themselves on Twitter is 10% of the number of accounts on Twitter {...} many languages are not included” (TC62).

The accuracy problem

Models can give information that is not completely exact, so caution needs to be taken during the interpretation and at the level of decision-makers: "You have to think that this model has a certain precision and it makes a certain number of errors {...} the place of declaration of the disease is not necessarily the place of infection. Therefore, if we try to predict the prevalence of a disease from environmental data, the fact that from this bad position, it may raise questions” (TC62).

Sensitiveness and information life cycle

When sensitive data are collected, they can be used for another purpose different from the first purpose, and this could impact the individuals concerned. An expert of geomatics working on foie gras producers stated: “The poultry farmers were directly impacted by the work we had produced {...} there are things that can be... political”. This means that any political decision made according to GPS localisation and disease risk may hamper the production capacity and the future of farmers (AT95).

At the level of the storage of all these data, a partner managing the MOOD prototype platform informs us that data are hosted in a cloud belonging to Amazon servers. The choice of Amazon as a supplier seems to be related to several advantages, such as the minimal risk of losing data and the high recognition of the good management of IT security risks by Amazon. Moreover, one data scientist did not have in mind at the time of our interview the particular EU legislation applicable to Amazon Cloud Computing services.

Possible solutions to mitigate ethical risk

During interviews, scientists specialising in machine learning and manipulating big data were still reluctant to consider the legal impact of the technologies they created. The legal aspect of AI was seen as an obstacle to research and innovation. Generally, two approaches exist in the MOOD project depending on who handles expertise: on the one hand, one of the experts has a philosophical background and privileges “soft law” or the ethical aspect of AI, while another expert assumes the legal aspect. Researchers were seeking support in their work; they asked for a “reminder of the tools that exist” and "ethical procedures regularly." Some needs are already identified by ethics experts themselves, such as the request for "defining procedures and strategies during global health emergencies" and a "code of good practice for manipulation of data” (ML50). Frequently Asked Questions (FAQ) on data and AI ethics in a small video presentation format or during webinar sessions can also be a solution. Researchers underlined "no legal safeguards to help us and to support us [...] " and " lack support, a help” and proposed a follow-up to mitigate it (TC62).

These results provide clues and highlight crucial subjects to be dealt with, such as data collection, confidentiality, understanding of the functioning of the automatic process linked to artificial intelligence and empowering health scientists in ethics knowledge related to big data and AI.

3.2 Quantitative

3.2.1 Survey

In total, sixteen people (out of 132; 12%) from different WPs and institutions responded to the questionnaire.

Regarding knowledge of ethics of data and AI, anonymization is perceived as sufficient (8/13; 62%); however, privacy by design (9/13; 70%) and encryption (10/13; 77%) are perceived as insufficient, and some lack knowledge on how it works. Property rights, re-identification and impact assessment were also perceived as mainly insufficient (8/13; 62% respectively) (Fig. 2).

Regarding knowledge of data protection and its legislation, 54% (7/13) of respondents had no or insufficient knowledge, while 46% of respondents (6/13) were considered to have sufficient knowledge on the topic.

Regarding machine learning and the "black box" (knowledge of how a given algorithm/model works), 61.5% (8/13) understood what was going on; however, 38.5% (5/13) of the respondents did not fully or not understand what was done automatically.

The data and databases received by the researchers mainly met the criteria for confidentiality (9/11, 81.8%), targeted the necessary information (8/11, 72.7%) and were considered non-discriminatory (7/11, 63.6%) (Fig. 3).

The majority of respondents (8/14; 57%) judged the data they manipulated with a lower level of sensitivity, followed by 4 researchers who judged the data they manipulated as highly sensitive 28.6%. Raw data was the most frequent format received by researchers (10/15, 67%), followed by pre-treated data (6/15; 40%) and completely anonymised data (6/15; 40%) (figure not shown).

The data were mainly anonymized by aggregation (8/11, 73%), taking into account geographical categories (11/12, 92%) socio-economic groups (2/12, 17%) or perceptions of the disease (2/12, 17%) (Fig. 4A and B). The remaining 46% (5/11) of respondents provided anonymised data with codes and identifiers (Fig. 4A).

Figure 4: Diagram of A) How researchers anonymize data and B) the parameters taken into account for the adaptation of data aggregation

Half of the respondents perceived the risk of deidentification in their data (7/14, 50%). A third of the respondents said they could not update or delete data from the database (2/6, 67%). Half of the respondents did not know who had access to their work (50%).

Access to the MOOD prototype platform requires a certain infrastructure (9/13, 69%), skills (9/13, 69%) and adequate equipment (8/13, 62%) (Fig. 5).

Concerning the knowledge of ethical principles and guidelines, the majority of researchers have read or are aware of at least the GDPR (General Data Protection Regulation) (14/16, 88%), some knew about Digital Health Ethics (WHO) (5/16, 31%) and EU ethics guidelines for trustworthy AI (3/16, 18,7%) and two people have not read any ethical documents on the subject (2/16, 13%) (figure not shown).

Regarding interactions with MOOD ethics experts, the majority of respondents (10/16, 63%) had already interacted with the experts; and 38% (6/16) of the respondents admitted never had done so (figure not shown).

Regarding the quality of exchanges with the Internal Ethics and Data Protection Board, the respondents rated equally i) accessibility, ii) understanding of what is requested, and iii) satisfaction with the feedback, with a score of 6 out of 10. Respondents were less satisfied with the interaction frequency (note 5 out of 10).

For ethical support, most of the respondents proposed a FAQ (19%), reminders of the main ethical issues (17%), and clear and common definitions (17%) (Fig. 6).

Several topics were suggested by respondents to be discussed in additional ethics sessions, such as support for the law and guidelines for data collection (15%, 2/14), avoidance of the risks of deidentification (15%, 2/14), and appropriate communication and dissemination of information and results (15%, 2/14) (Fig. 7).

Governance of data and ethical risks identified

All the systems that involve the collection, curation and analysis of data imply engaging with data governance that encompasses a multitude of social systems (19). The form of data governance privileged by the MOOD consortium promoted a data-sharing approach, but the reality showed that it is less easy than it is supposed to be because it implied adherence to a set of rules and compliance with the existing EU legal systems and non-EU countries. For example, data access to the EFSA platform required an access agreement and reuse agreement through motivated requests. The network, nodes, data flow and connectivity also show a crucial challenge posed by big data analytics: the interoperability of systems that led to several issues such as standardisation related to a complex infrastructure.

The mapping shows that several types of ethical risks can appear at different levels of the path taken by the data.

The issue of privacy is a main challenge, as announced by an ethics consultant in an interview and confirmed by the literature. Owing to the cross-referencing of data, deidentification is always possible (20). Several public health scholars have shown how data brokers commodify personal data for profit purposes (21). As one expert of ethics assesses, merging data collected by researchers with other existing data can lead to unexpected problems or unwanted harm.

Machine learning and its algorithms are important in predictions and real-time analysis. Nevertheless, they could reproduce stereotypes included in a set of nonofficial sources and lead to problems of discrimination or stigmatisation of groups depending on their race or gender. The COVID-19 pandemic generated information with negative views about Chinese ethnic groups, as Wuhan, which is situated in central China, was the first area where early cases were reported, and China was considered the hot zone where the virus and death originated. Owing to the selective nature of data and data-driven techniques, an increasing number of studies have shown that some correlations could serve as proxies for unknown or protected categories that were deliberately unrecorded, techniques that could lead to gender inequities and public health services provided to the detriment of one sexual group (22). Some disease intelligence tools tailored for extracting epidemiological events through media news for syndromic surveillance purposes and disease profiling activities often ignore gender-related trends. Scholars currently and historically investigate how ignoring differences amplifies discrimination (22).

The black box approach that leads to opaque algorithms limits the fundamental freedom and reliability of AI systems related to FAIR principles that are critical for decision-makers; the lack of transparency could impact decisions based on the outputs of AI systems. AI transparency and explainability are the six core ethical principles of the United Nations (23).

Data from social media sources are anonymised. Moreover, information deleted by users from online social media sources remains in the researcher’s database (ex, a tool that extracts unofficial data for syndromic surveillance), which can render datasets obsolete and may pose a critical issue that confronts the right to be forgotten to actual means to do so. The GDPR obliges the legal person who commits the data processing to stop processing inaccurate or incomplete, according to Twitter's terms and conditions, a compliant user API should remove critical data from the database built via Twitter’s API (25). It is not easy to do so when the process is ongoing. In addition, Twitter data access is provided through an API that makes the data publicly available to anyone. An important ethical issue remains when informed consent is not requested for public health surveillance in digital publics that provide research outputs. We can also see the possible existence of sampling biases validated by interviews for representativeness at the level of Twitter users who are not representative of the general population or keyword searches used by the algorithm. A misinterpretation of the content of the tweets is also possible without considering the context of the identified keywords, as mentioned by a project partner during a discussion. An ironic tweet saying “to prefer an Ebola pandemic rather than losing soccer games” was wrongly collected by an algorithm. Similarly, for the societal context, events or restrictions can influence the content and latest trends. These observations confirm the importance of semi-automatisation and the requirement of human supervision during machine learning and text extraction, especially during the pandemic when European disease surveillance agencies lack human resources (24). Unfortunately, Twitter data first used as experimental ground were no longer used in PADI-Web but remained a valuable source of data for ProMED. The lack of human resources to continue monitoring disease trends and strict ethical risks (privacy concerns) may lead European researchers and data scientists to talk about mining social media sources.

The Euro-American tradition of research driven by industry with enormous data centres and infrastructure located in places such as Oxford, California, or Boston is internationally well-known. The approach allows the imposition of criteria that are universally recognised such as the English language used for the majority of algorithm keywords in detrimental vernacular languages (eastern European or African languages). The language orientation of surveillance systems tells us that data are highly driven by politics (15). The vernacular indifference of machine learning may amplify discrimination.

During informal discussions, the expression “radioactive data'' appears several times when we listen to experts and it is immediately related to personal data such as GPS, some confusion has been voiced among researchers who are not perfectly aware of what personal data truly embraces. Personal data are not dangerous by themselves; it is the hands that manipulate data and the objectives and interests, whether commercial, political, or repressive, under the police investigation that could impact people who could for example lose their freedom or have financial losses. Given that the use of big data and AI can be a “double-edged sword” (26), researchers should think critically about what the larger societal consequences of their intervention might be. Building ethical sociotechnical systems requires the shifting of AI ecosystems toward equitable and accountable AI. This became a major objective for the European Commission; researchers were required to seek the consent of users, avoid privacy intrusion, and use data minimisation (17, 27). Knowledge of these risks by researchers is crucial.

Knowledge and role of researchers

The approach of ethics expertise that confronts comprehensive moral guidance versus enforcement of legal compliance (depending on expert background), a kind of “soft” versus “hard” ethics (28) has consequences for some researchers' readiness to cooperate in thinking and tackling ethical issues raised by their tools and sociotechnical innovation.

The answers collected throughout the survey show that the respondents’ level of ethical knowledge varies according to their field of expertise, whether they are researchers or data scientists. The respondents evaluated their knowledge as good for anonymization (Fig. 2), somewhat opposed to how they received and anonymised the data. A few parameters, although ethically important, are considered to adapt the aggregation of the data (Fig. 4A and 4B). Additionally, we have seen above that not all respondents have read or are aware of the GDPR. As partners outside the European Union may not have the same data protection legislation, it is important in the MOOD project to establish common standards for working internationally, nevertheless, partners in Eastern Europe are invited to follow their regulations.

Conditions of access to the MOOD platform may cause a “digital divide” mainly for partners in the southern region because it required special conditions (skills, strategies, instrument information and good internet connections) are required despite the "open source" perspective adopted by the project that makes models and codes available.

Big data is inherently voluminous; the storage allowing them to be available must be able to follow. As seen during interviews and on the MOOD data flow map, all project data are hosted by AWS (Amazon Web Services) servers, an American company. One scientist suggests that the physical location of the server is only sometimes known. The “CLOUD Act” (Clarifying Lawful Overseas Use of Data Act), a 2018 U.S. federal law “expands the geographical scope of possible requests by the U.S. government that can access data on servers, regardless of their location”. This law tends to go against the GDPR even though an international agreement remains mandatory “for a jurisdiction or an authority resulting from an administration to transfer or disclose personal data”.

In the event of disagreement, a U.S. court may provide a warrant if the latter is convinced that the public interest is threatened. In addition, the European legislation has several requirements related to data storage. The cloud service offering must be a public and multi-public provider, hybrid, and energy-efficient, which means lowering the carbon footprint (29).

One of the respondents considered environmental data to be without ethical risks and, therefore, without the need for aggregation (see Fig. 4A). While reflecting on the future of a platform integrating these data, we may assume that considering certain environmental parameters as disease emergence factors could not harm the biodiversity of a place. Indeed, if one wishes to reduce the risk of emergence in the future, this could imply lowering or eliminating certain environmental factors by reducing or shaping part of the biodiversity. When we reflect on the "One Health" approach, it may have an impact on the environment or animals that deserve to be carefully considered by researchers and decision-makers.

Ethics help support

The current advice provided by ethics advisors does not seem to be enough for several researchers. This can also be seen in evaluating the ethical advice in the questionnaire and the averages attributed to each criterion. Figure 6 shows the types of assistance expected in addition to an ethical code of conduct as support. However, 57% of the respondents (9/16) considered that they did not need an ethical code of conduct to perform their job, or possibly in another form.

Study limitations

Several internal documents used to create the data flow map have been updated since then, making some information obsolete as the project changes over time. Therefore, the actor-network is not exhaustive because it was constructed at only one point in time, and this should be considered when interpreting the results.

Regarding the interviews, the timeline for implementing and analysing the data was short. Despite the population target (MOOD partners N = 132), the sample was small (16/132), we obtained few answers, and people had only 22 days to answer, therefore, the results were not statistically significant. Moreover, implementing mixed methods helped to balance this issue, and the methodology approach was intrinsically inductive.

For a more balanced and richer overview of ethical concerns in the governance of data and the use of algorithms in digital health surveillance, it was important to consider the views of end users of the MOOD prototype platform, which were not identified when our study started. Ethics issues related to the use of the prototype platform may be assessed in future evaluations of the program that must incorporate the coming AI act of the EU and the installation of a paywall for Twitter APIs that will considerably change the political economy of open source in academic research. Ethical issues also require being more technically oriented in providing concrete answers to tangible problems.

Our case study illuminated complex issues related to the management and perception of digital health data ethics in a One Health project. First, data as an object is complicated by its statute when operating within a big data analytical infrastructure. As we can see, when the data are collected in silos, the data coming from the environment, public health records and animals are stored in different infrastructures owned by different organizations, using different standards. Therefore, any governance of data needs to challenge the issue of interoperability.

Second, considering that safety, security, and the right to privacy should be addressed by AI actors, the MOOD project established a data protection framework that sheds light on how to protect and promote the Big Data lifecycle. Ethic advisors came to support researchers in being compliant with EU directives in terms of data science and AI.

However, the data considered by researchers themselves as “not verified”, “not validated” and unstructured” include new sources such as social media whose values are changing as they accede to scientific evidence because the methodology that is used deserves to be scrutinised in epistemology, law and ethics implications to services provided to society.

It seems that the proliferation of principles and guidelines in the ethics of data and AI is not always easy when considering them alone, for researchers and data scientists working in the One Health program, the request is in favour of learning by doing, ethics advisors that promote critical thinking and strong support and tools for remembering the ethical principles that drive data processing and analytics.

In the end, the data analysis infrastructure that operates under open science rules needs to follow the FAIR principles, remain accessible and be sustainable in the future.

Ethics approval

Ethics approval for this work came from the VetSupAgro ethics committee Ref decision: 2173 regarding task 6.4 (Outcome assessment for impact evaluation) of the MOOD H2020 project.

Consent for publication

Not applicable.

Availability of data and materials

The datasets generated and/or analysed during the current study are available but anonymized. You may ask them from the corresponding author on reasonable request.

Competing interests

None

Informed consent:

Oral informed consent to participate in the work was obtained from project partners interviewed for the study according to the Grant Agreement

Funding

This study was funded by the “Monitoring outbreak events for disease surveillance in a data science context" (MOOD) project from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 874850 (https://mood-h2020.eu/) and is catalogued as MOOD 0088. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Authors contribution

EA is a member of the Ethics Board, she explained, transferred internal documents for review and revision of the manuscript

LM: collected, analysed and wrote the quantitative section of the manuscript

ST: participated in the methodology and reviewed the paper critically

OT: designed the study, chose the methods, analysed the data and wrote the manuscript

Acknowledgements

We thank all the MOOD partners who took the time to answer our questions. We are grateful to Pierrine Didier (VetAgroSup/INRAE, France) and Henok Tegegne (ANSES, France) for their input in the data mapping.

Brownstein JS, Freifeld CC, Madoff LC. Digital disease detection—harnessing the Web for public health surveillance. N Engl J Med. 2009;360(21):2153.
Yousefinaghani S, Dara R, Poljak Z, Bernardo TM, Sharif S. The Assessment of Twitter’s Potential for Outbreak Detection: Avian Influenza Case Study. Sci Rep 3 déc. 2019;9(1):18147.
Jordan S, Hovet S, Fung I, Liang H, Fu KW, Tse Z. UsingTwitterforPublicHealthSurveillancefromMonitoringandPredictiontoPublicResponse.Data.29déc2018;4(1):6.
Hartley D, Nelson N, Walters R, Arthur R, Yangarber R, Madoff L. The landscape of international event-based biosurveillance. Emerg Health Threats J avril. 2010;3(1):7096.
Paquet C, Coulombier D, Kaiser R, Ciotti M. Epidemic intelligence: a new framework for strengthening disease surveillance in Europe. Euro 1 déc. 2006;11(12):5–6.
Signorini A, Segre AM, Polgreen PM. TheUseofTwittertoTrackLevelsofDiseaseActivityandPublicConcernintheU.S.duringtheInfluenzaAH1N1Pandemic.GalvaniAP,éditeur.PLoSONE.4mai2011;6(5):e19467.
Arsevska E, Valentin S, Rabatel J, de Goër de Hervé J, Falala S, Lancelot R. etal.WebmonitoringofemerginganimalinfectiousdiseasesintegratedintheFrenchAnimalHealthEpidemicIntelligenceSystem.DóreaFC,éditeur.PLOSONE.3août2018;13(8):e0199960.
Hou Z, Du F, Jiang H, Zhou X, Lin L. Assessment of public attention, risk perception, emotional and behavioural responses to the COVID–19 outbreak: social media surveillance in China. Risk Percept Emot Behav Responses COVID–19 Outbreak Soc Media Surveill China 362020.2020.
Espinosa L, Wijermans A, Orchard F, Höhle M, Czernichow T, Coletti P. etal.Epitweetr:EarlywarningofpublichealththreatsusingTwitterdata.Eurosurveillance[Internet].29sept2022[cité2oct2023];27(39).Disponiblesur:https://www.eurosurveillance.org/content/10.2807/1560–7917.ES.2022.27.39.2200177.
Coulombier D. Epidemic intelligence in the European Union: strengthening the ties. Euro 7 févr. 2008;13(6):1–2.
Borda A, Molnar A, Neesham C, Kostkova P. Ethical Issues in AI-Enabled Disease Surveillance: Perspectives from Global Health. Appl Sci janv. 2022;12(8):3890.
Awrahman BJ, Aziz Fatah C, Hamaamin MY. A Review of the Role and Challenges of Big Data in Healthcare Informatics and Analytics. Abdelaziz M, éditeur. Comput Intell Neurosci 29 sept. 2022;2022:1–10.
European Parliament. Civil liability regime for artificial intelligence [Internet].2020/2014(INL). oct 20, 2020.Disponible sur:https://www.europarl.europa.eu/doceo/document/TA–9-2020-0276_EN.html.
Boyd D, Crawford K. Six Provocations for Big Data. SSRN Electron J [Internet].2011[cité 2 oct 2023]; Disponible sur:http://www.ssrn.com/abstract=1926431.
Leonelli S. La recherche scientifique à l’ère des big data: cinq façons dont les big data nuisent à la science et comment la sauver. Milan: Editions Mimésis; 2019. p. 118. (Philosophie).
Andrejevic M, Gates K. Big Data Surveillance: Introduction. Surveill Soc 9 mai. 2014;12(2):185–96.
EuropeanCommission. Directorate-General for Communications Networks, Content and Technology. Ethics guidelines for trustworthy AI [Internet]. Brussels: Publications Office; 2019. Disponible sur. https://data.europa.eu/doi/10.2759/346720.
Kondylakis H, Kalokyri V, Sfakianakis S, Marias K, Tsiknakis M, Jimenez-Pastor A. Data infrastructures for AI in medical imaging: a report on the experiences of five EU projects. Eur Radiol Exp 8 mai. 2023;7(1):20.
Beaulieu A, Leonelli S. Data and society: a critical introduction. London: Sage Publications Ltd; 2022. p. 246.
Vayena E, Salathé M, Madoff LC, Brownstein JS. Ethical Challenges of Big Data in Public Health. PLOS Comput Biol févr. 2015;11(2):e1003904.
Ebeling MFE. Healthcare and Big Data: Digital Specters and Phantom Objects. Palgrave Macmillan; 2016. p. 181.
Chun WHK. Discriminating Data: Correlation, Neighborhoods, and the New Politics of Recognition. Cambridge, MA, USA: MIT Press; 2021. p. 344.
UNESCO.UNESCO’s Recommendation on the Ethics of Artificial Intelligence:keyfacts[Internet].UNESCO. ;2023[cité31août2023]p.19.Disponiblesur:https://unesdoc.unesco.org/ark:/48223/pf0000381137/PDF/381137eng.pdf.multi.
Dub T, Mäkelä H, Van Kleef E, Leblond A, Mercier A, Hénaux V. Epidemic intelligence activities among national public and animal health agencies: a European cross-sectional study. BMC Public Health 4 août. 2023;23(1):1488.
Uršič H. TheRighttobeForgottenortheDutytobeRemembered?Twitterdatareuseandimplicationsforuserprivacy[Internet].CouncilforBigData,Ethics,andSociety.CouncilforBigData,Ethics,andSociety;2016[cité30août2023].Disponiblesur:https://bdes.datasociety.net/council-output/the-right-to-be-forgotten-or-the-duty-to-be-remembered-twitter-data-reuse-and-implications-for-user-privacy/.
Lin L, Hou Z. Combat COVID–19 with artificial intelligence and big data. J Travel Med.2020.
Mello MM, Wang CJ. Ethics and governance for digital disease surveillance. Sci 29 mai. 2020;368(6494):951–4.
Floridi L. Soft ethics, the governance of the digital and the General Data Protection Regulation. Philos Trans R Soc Math Phys Eng Sci. 2018;15(2133):20180081.
European Commission. The European Commission Cloud Strategy [Internet].2019[cité 30 août 2023]p.28.Disponible sur:https://commission.europa.eu/publications/european-commission-cloud-strategy_en.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Data governance and ethics in digital health surveillance for emerging infectious diseases

Status:

Version 1

Abstract

Figures

Key Points

1. Introduction

1) Description of the project’s dataflow;

2. Materials and methods

2.1 Overview of the MOOD project

2.2 Qualitative analysis

2.2.1 Mapping

2.2.2 Semi-structured interviews

2.3 Quantitative analysis

2.3.1 Survey

3. Results

3.1 Qualitative

3.1.1 Mapping

3.1.2 Semi-structured interviews

3.2 Quantitative

3.2.1 Survey

4. Discussion

5. Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1