Navigating ethics in AI-enabled health research: insights from four case studies

Introduction

Overview

Artificial intelligence (AI) is increasingly embedded in health systems worldwide, shaping decisions related to diagnosis, treatment, surveillance, resource allocation, and health policy. AI offers tremendous potential to improve efficiency, reduce errors, lower healthcare costs, expand access to care, and enhance population health outcomes (1,2). AI is now used across clinical medicine and public health for diagnostic support, risk prediction, workflow optimization, patient self-management, and much more (3,4). By continuously monitoring and analyzing patient data, AI can also be used to alert patients and/or providers to early signs of disease or complications, enabling proactive intervention and preventive measures (3). The use of predictive data analysis can also anticipate potential health issues, as AI systems can identify patterns and anomalies in data that might be challenging for human professionals to detect (3). AI has shown high accuracy in detecting cancer (5-7), cardiovascular and neurological events (8-10), infectious diseases (11-13), retinal disease (14), mental health disorders (15-17), and many more conditions. In health research, AI has also been used to develop early warning and response systems for epidemics, natural disasters, and crop failures (18-24).

However, numerous complex ethical issues are associated with the use of AI for health research, which may impact healthcare providers, researchers, patients, governments, and policymakers. Decisions about how AI systems are designed, deployed, and governed can have real-world consequences for providers, healthcare systems, and patients, particularly for underserved populations. Robust governance and policies are needed to ensure responsible AI use (25). Yet the current global regulatory landscape is characterized by significant variation in frameworks across jurisdictions, making governance challenging (26-28). As seen in Table 1, four key ethical issues will be reviewed in this collection of case studies: (I) privacy, security, and data management; (II) autonomy and consent; (III) transparency and accountability; and (IV) fairness and bias. While these four domains do not represent an exhaustive list of ethical considerations related to AI, they are particularly relevant to understanding ethical issues that may arise in health research and human subjects oversight. Together, they provide a practical foundation for examining how Institutional Review Board/Research Ethics Committee (IRB/REC) members might identify, assess, and navigate such issues in health research.

Ethical issues

Privacy, security and data management

One of the most common concerns is privacy and data security (2-37). Data security is a significant concern because AI tools are dependent on large-scale health datasets and often contain vast amounts of sensitive patient data. Protecting data from unauthorized access, breaches, or malicious attacks is essential (4,28,31,38-40). In the health research context, “privacy” usually includes:

• Limiting use to appropriate purposes (research aims, not unrelated secondary uses);
• Minimizing identifiability (de-identification/pseudonymization; careful handling of linkages);
• Controlling access and disclosure (who can see what, under what conditions);
• Accounting for AI-specific risks (e.g., model inversion/membership inference, re-identification via data linkage, leakage through outputs).

A recent innovation is Federated Learning that enables healthcare institutions to send algorithmic updates to a central server without the need for sensitive patient data to leave the hospital. However, Federated Learning is not immune to algorithmic bias and the “black box” problem.

Ethical data management begins with informed consent that explains how data will be used, reused, and potentially linked with other datasets. It also requires robust de-identification practices, secure data infrastructures, and compliance with relevant regulations. Data transparency, information on how data are collected, stored, and used in the development of an AI system, is also critical (41). Further, data sovereignty is another dimension. It is the idea that data are subject to the laws and governance structures of the nation in which they are collected is a crucial dimension of data protection. However, regulatory environments differ significantly across jurisdictions, which poses challenges for standardization of guidelines (42).

Autonomy and consent

Researchers have raised the importance of maintaining patient autonomy and ensuring informed consent in AI-assisted health research. For example, many wearable users have inadequate understanding of the terms of their consent and how their personal data are used (43). Obtaining informed consent upholds patient autonomy by reaffirming patient control over their own health data (44). The integration of AI into health research may lead to situations where decision-making is either transferred to or shared with machines. Upholding the principle of autonomy necessitates that AI-assisted health research must not compromise human autonomy. Respecting autonomy also encompasses the implementation of appropriate ethical oversight (3).

Transparency

Trustworthiness and the lack of transparency are also major concerns for researchers, clinicians, and patients (4,30,32,33,44). Transparency in this context can be seen as information about the mechanisms behind why algorithms produce specific outputs (45). Relatedly, the term “explicability” may be used to represent transparency, as it refers to the principle that AI decisions should be explainable to those affected to the extent possible (45). AI systems often operate as “black boxes” with little information on how decisions are reached (28,31,46). Transparency requires that AI systems provide adequate information about decision-making processes so that various stakeholders such as clinicians and research participants can understand the performance and limitations of the system, in addition to interpreting decisions (41,45). For example, to trust AI-based data, researchers must have confidence that the insights generated are from representative data that cover relevant demographic and clinical variations (44). Transparency is crucial because it exposes the inner workings of the AI system, which can help to ensure ethical research protocols (45,47). Explainable AI is a promising new intervention to address transparency and explainability concerns. It allows opening the “black box” by translating algorithms to terms that can be understood by humans. However, current models are still imperfect and are not free from trust and accountability concerns.

Fairness and bias

Data quality and bias in AI algorithms are also major concerns. Algorithms may mirror human biases in decision making (48). Bias can arise from biased training data or decision-making processes, potentially exacerbating disparities in health care outcomes (3,4,28,30,31,40,44,48-52). For example, underrepresentation of marginalized populations in training datasets can create inherent biases in the models, thereby perpetuating discrimination. Often issues of fairness are only apparent after examining the results of the model, such as what the model is optimizing for and who is missing from the data (31).

Objective

This collection of original case studies is a practical tool to facilitate discussion and critical reflection on the key ethical dimensions of AI in health research. The goal is to encourage critical reflection on ethical clarity, completeness, and practical review challenges, prompting dialogue among diverse stakeholders, including students, clinicians, public health professionals, policymakers, and researchers.

Establishing interdisciplinary ethics committees or review boards to evaluate the various ethical dilemmas in AI for health research is increasingly necessary (28). More training is required in order to equip these members with the skills necessary to assess the protocols for ethical research utilizing AI (25,53,54). By engaging with realistic scenarios of ethical challenges in AI for health research from different settings, learners can explore how ethical principles interact with social, cultural, legal, and political contexts. The cases can also highlight important tensions in values, such as privacy vs. public benefit and efficiency vs. patient autonomy. The four cases included in this paper are also intended to provide a platform for discussion about appropriate governance responses (e.g., auditing, oversight committees, escalation pathways, monitoring plans, community engagement, etc.). Discussion questions for each case study are included to facilitate dialogue (see Tables 2-5).

The four ethical domains presented in these cases were chosen based on common issues that a researcher or member of an IRB/REC might encounter in health research. The co-authors highlighted four distinct scenarios that would be timely and relevant and that would specifically illustrate the four ethical domains. Fictional names were used for people, organizations, and places. We pilot tested the cases at a two-hour, in-person workshop that was attended by global health students and professionals at a conference in Malaysia in late 2025. After receiving feedback from the participants, we made significant revisions to the cases.

Case presentation

Case 1: the unintended breadcrumbs

Introduction

Dementia is one of the most rapidly growing health challenges of the twenty-first century, with diagnosis rates expected to reach 139 million by 2050 (55). Alzheimer’s disease, which accounts for 60–80% of these cases, is a progressive neurodegenerative disorder in which pathological brain changes often begin decades before clinical symptoms appear, resulting in a long preclinical phase during which the disease is active but without observable symptoms (56). Diagnosis typically happens only after a significant and observable cognitive decline. This delay inversely closes the window for early interventions that could preserve a patient’s quality of life (55). One persistent difficulty in diagnosing lies in the limitations of current standard assessments. This is because traditional cognitive tests, which are often conducted during brief clinical appointments, capture only a snapshot of a patient’s condition and can fall short of detecting subtle changes across diverse populations (57,58).

AI and the “data-driven” diagnosis

In recent years, AI, particularly predictive modelling and deep learning (DL), has been promoted as a means of addressing the aforementioned diagnostic limitations (59). The application of AI to this condition is diverse. For instance, neuroimaging, DL algorithms (specifically convolutional neural networks) are now used to analyze magnetic resonance imaging (MRI) and positron emission tomography (PET) scans to detect minute patterns of brain atrophy invisible to the human eye (60). Similarly, natural language processing (NLP) tools have been developed to analyze speech acoustics to identify subtle pauses or vocabulary changes in a patient’s voice that may signal early cognitive impairment (61). See Appendix 1 for definition of these terms.

However, while imaging and speech analysis show promise, they often rely on patient active participation. Hence, a more scalable avenue is to utilize data already generated through routine care. By analyzing vast repositories of archived electronic health records (EHRs), AI systems can identify non-linear patterns that human clinicians might miss, such as subtle shifts in vocabulary within clinical notes and complex correlations between seemingly unrelated prescriptions (62).

The transition of these models from research to reality is fraught with challenges. In particular, these algorithms depend heavily on the diversity and volume of the datasets they consume to function effectively. This dependence creates a profound dilemma between the clinical need for medical history and the ethical imperative to protect patient privacy (63). As AI systems feed on detailed data, including unstructured breadcrumbs of a patient’s life found in clinician notes, they raise questions about consent, anonymity, and the risk of re-identification in the age of big data (38).

Case scenario

Dr. Samuel Rivera, a neurologist and principal investigator at Manila City Medical University (MCMU), has secured a large grant to develop a machine learning model capable of predicting early-onset dementia up to five years before clinical diagnosis. The long-term vision is to integrate the model into routine primary care screenings, enabling earlier interventions, lifestyle adjustments, and participation in clinical trials. This could have positive impacts in terms of reducing healthcare costs, as well as advancing health equity by detecting underdiagnosed cases in marginalized populations. He plans to build an AI model capable of predicting early-onset dementia using ten years of archival data. Specifically, it will make use of NLP in order to analyze human language computationally (64). The NLP component of the model will scan unstructured clinical notes for subtle linguistic and behavioral markers for cognitive decline, such as shrinking in vocabulary, instances of misunderstanding instructions, and casual mentions of struggling with regular tasks. By synthesizing these clues through an algorithm, a predictive risk score can be generated. The dataset includes 250,000 patient records with laboratory results, medical history, and unstructured clinical notes, collected from 2010 to 2020 from patients aged 60 or older who received care at the MCMU’s affiliated healthcare system. To prepare the data for the study, all direct identifiers, such as names, addresses, phone numbers and full dates (except year), were removed in accordance with United States’ federal standards for protecting health information (called the Health Insurance Portability and Accountability Act, or HIPAA). The HIPAA Safe Harbor standard is a legally recognized method of data de-identification under the United States’ HIPAA that mandates the removal of 18 specific types of personal identifiers to ensure that health-related data can be safely used for secondary research without risking patient re-identification. Through the rigorous de-identification process, the research team believes that the data is safe for analysis without requiring individual consent, which would be impossible given that many patients are either disease-free or uncontactable. Because the records span a whole decade, thousands of patients have since relocated, changed hospitals, replaced their contact information, or passed away. Furthermore, these patients likely signed the standard hospital intake forms agreeing that their de-identified data could be used for broad research or quality improvement purposes. However, none of them explicitly consented to having their ten years of intimate medical narratives be used to be analyzed by a predictive AI algorithm. In the end, when the protocol for the study was submitted, the IRB determined that because the dataset no longer constituted “human subjects research”, and therefore individual consent was not required.

However, the assumption of anonymity was challenged during a meeting when Sarah Malik, the lead data scientist, projected a random sample of the de-identified clinical notes to demonstrate preprocessing steps for NLP. While the direct identifiers were absent, the text included detailed and vivid information about the patients, including:

• “Retired pediatric neurosurgeon who practiced at St. Agatha’s Children’s Hospital for 35 years”.
• “Diagnosed with Wilson’s disease at age 42; one of three known cases in the county”.
• “Primary caregiver for spouse, former mayor of Brookdale”.

Sarah argues that these “breadcrumbs”—rare diseases, specific job titles, and distinct family histories—create a risk of re-identification. Further, she questions whether regulatory compliance inherently means the principles of ethical research are being upheld. The team now faced a critical dilemma: whether to remove these details and potentially hamper the mode’s ability to detect subtle cognitive decline, or to proceed with the data as is, relying on technical compliance with HIPAA standards to justify the ethical risks. For example, unstructured notes may contain subtle signals of early cognitive decline. For example, behavioral descriptions might be embedded in the text or clinicians may note things like “patient reports being increasingly forgetful but managing well”. Dr. Rivera argues that removing information about life events or job titles might reduce re-identification risk but it could also degrade predictive performance. Further, he worried that it could potentially bias the model toward more stereotypical presentations of dementia.

Main ethical concerns

Privacy, security and data management

First, the central tension in this case lies in whether adherence to legal frameworks is sufficient to guarantee ethical protection in the age of big data. Is de-identified data truly anonymized when an AI can cross-reference it with large datasets that could re-identify individuals? While the research team followed HIPAA standards by removing direct identifiers, this checklist approach is often insufficient for high-dimensional datasets like clinical notes or genomic profiles. The dilemma here is the mosaic effect, in which disparate pieces of innocuous data, such as rare diagnosis combined with a specific occupation mentioned in the clinician’s note, can be combined to reveal a person’s identity (65). If the researchers proceed with the raw notes to maximize the system’s predictive capacity, they risk a privacy breach that violates the core principle of non-maleficence. In the same manner, if they aggressively remove data, they may strip away the minute signals the system needs to detect early dementia, rendering the study useless.

The trade-off between algorithmic utility (ensuring the model works for complex, real-world patients) and the potential stigmatization if the patient is re-identified or if the model parameters are eventually published is another important dilemma. The predictive models, once published, could be misused by insurance companies and employers to screen out high-risk individuals before they even show symptoms. This is an example of a dual-use dilemma: having the same algorithm designed to help clinicians can be weaponized to deny coverage. If the researchers publish their model parameters to advance science, they may lose control over how the system is used in the real world, potentially violating the principle of justice by disproportionately harming vulnerable populations flagged as high risk by the system (66). This begs the question: how to move forward with dissemination of the results to both advance science and help the public?

Autonomy and consent

This case also raises questions regarding the moral authority of the researchers to decide on the risks on behalf of the uncontactable and disease patients and the justification of waiving consent when the secondary use of data involves advanced predictive modelling of a medical condition. Dr. Rivera argued that obtaining consent from 250,000 retrospective patients is impractical, justifying that many patients were either deceased or uncontactable. However, this creates an ethical deficit regarding patient autonomy. After all, patients who visited the hospital in 2010–2020 did not foresee, nor agree, that their intimate medical notes would be used to train a predictive AI model. This broad consent approach raises concern about whether the researchers are effectively stripping patients of their right to control their data in the name of efficiency (67).

Guidance

Privacy, security and data management

It is important to bear in mind that ethical challenges are context-sensitive, as regulations tend to be reactive to specific problems. Given the age of big data and AI, which has ushered in a new research climate, a top-down adoption of the HIPAA Safe Harbor standard might be insufficient. To address the gap, the team should move beyond the standard by performing manual redaction and employing advanced privacy-preserving technologies, such as differential privacy. This method adds calculated statistical noise to the dataset, allowing the AI system to learn general patterns of dementia across the population without reproducing the specific data of any individual (68).

However, implementing differential privacy in clinical research may involve a delicate tradeoff between privacy budget and model utility. This tension is more explicit in small, high-dimensional clinical data sets such as unstructured EHRs, where excessive statistical noise can erase critical and implicit clinical signals necessary for accurate algorithmic predictions. Thus, rather than treating it as a frictionless plug-and-play solution, investigators using differential privacy must carefully and routinely calibrate and audit its privacy-utility curve. Additionally, for unstructured clinical notes, the team should use algorithms equipped with named entity recognition to automatically detect and redact indirect identifiers, such as professions or locations, before entering the data into the AI system.

Ethical research necessitates recognizing that social harms are as important as scientific harms. To mitigate the social harm the system causes through discrimination and stigmatization, the research team should adopt a controlled release strategy rather than immediately open-sourcing the raw model. This would mean publishing the methodology while keeping the trained AI system in a restricted-access environment, available only for scientific research. Furthermore, any publication related to the AI system should include model information that explicitly details the model’s limitations, intended use, and out-of-scope uses (69).

Autonomy and consent

Consent cannot always be interpreted as a blanket in the absence of the patient. Given that individual consent is impossible, the solution would depend on robust data governance. Hence, the institution should use Trusted Research Environments (TREs) so that data never leaves the hospital server; researchers send their AI systems in, and only aggregated, vetted results come out (70). TREs serve as highly secure computing platforms that provide authorized researchers with remote access to de-identified datasets without downloading or copying the original files. This can ensure that, even if the AI system finds a pattern, the raw patient data is never exposed to less-secure servers. Furthermore, beyond technical systems, the establishment of a Community Data Trust, a board composed of patient advocates and ethicists, can help provide oversight in ensuring that specific queries align with the community values, effectively providing social license to operate instead of the absence of individual consent.

Case 2: AI-driven cardiovascular risk prediction trial: a multinational academic study

Introduction

Cardiovascular disease represents the leading cause of death globally, necessitating accurate risk prediction and diagnostic tools to guide clinical decision-making and improve patient outcomes (71). While advances in prevention, diagnosis, and treatment have significantly reduced mortality in many high-income countries, these gains have not been shared equitably. Low- and middle-income countries (LMICs) still suffer from undiagnosed disease, delayed care, and preventable death. These disparities are not accidental. They are shaped by structural constraints, including limited healthcare resources, fragmented health information systems, and risk profiles that differ from those of the populations on which many clinical tools are based.

One persistent difficulty lies in how cardiovascular risk itself is assessed. Commonly used risk prediction models are largely derived from data drawn from relatively homogeneous populations and well-resourced healthcare settings. When applied elsewhere, their accuracy and relevance are often assumed rather than demonstrated (72). Improving early risk identification in ways that are clinically useful, socially credible, and ethically defensible therefore remains a central challenge in global cardiovascular health.

AI and cardiovascular risk prediction

In recent years, AI, particularly machine-learning approaches, has been promoted as a means of addressing some of the limitations of current cardiovascular risk prediction models (73,74). By combining large and heterogeneous data sources, ranging from EHRs and laboratory tests to imaging, genomics, and data from wearable devices, AI systems can produce probabilistic risk estimates that extend beyond the scope of conventional models (74). Proponents argue that such systems allow for more individualized and timely clinical decision-making by identifying complex patterns in high-dimensional clinical data that traditional diagnostics may overlook (73).

Empirical studies and systematic reviews have reported strong predictive performance for machine-learning-based cardiovascular risk models, particularly in controlled research environments (73,74). However, while “machine learning (ML) and deep learning (DL) approaches have shown increasing promise for heart disease prediction… comprehensive evidence regarding their effectiveness compared to traditional clinical methods, implementation challenges, and real-world deployment readiness remains fragmented” (71). At the same time, these systems raise familiar ethical and governance concerns. AI models can reproduce biases embedded in their training data, perform poorly when deployed in new populations, and operate in ways that are unclear to both clinicians and patients (75-77). Their integration into clinical practice may also subtly shift professional responsibility, alter patient expectations, and blur lines of accountability within healthcare institutions (77).

These concerns become more critical in multinational research settings. Differences among nations in terms of healthcare infrastructure, regulatory capacity, data quality, and patient familiarity with digital technologies can complicate both the evaluation of AI systems and the ethical oversight of their use (77,78). This case explores these issues through a multinational clinical trial of an AI-based cardiovascular risk prediction tool.

Case scenario

The proposed research study is a multicenter randomized controlled trial that will evaluate an AI-driven cardiovascular risk prediction model. This model was developed at the African Global Health Institute and funded by a private entity that is based in a country in Western Europe. The AI model generates individualized estimates of five-year risk for major cardiovascular events, including heart attack, stroke, and sudden cardiac death. Predictions are based on routinely collected clinical data, such as demographic characteristics, laboratory values, imaging findings, and self-reported lifestyle factors. The data are automatically integrated into the EHR that will guide patient care decisions. To train the AI model, developers used clinical data from thousands of European patients who have suffered from cardiovascular diseases, mostly adult white males.

The trial will be conducted at three sites representing distinct economic and healthcare contexts: Toronto Cardiac Disease Center (Canada), a high-income setting; Rio de Janeiro University Hospital (Brazil), a middle-income setting; and Nairobi Regional Heart Hospital (Kenya), a low-income setting. Each site will enroll approximately 5,000 adults aged 40–75 years with at least one established cardiovascular risk factor, such as hypertension, diabetes, or a history of smoking.

Research participants will be randomized into two groups. Group 1 will receive AI-supported care, wherein clinicians will have access to AI-generated risk scores during consultations. Participants in this group will not be informed that the risk scores are AI-generated. Meanwhile, Group 2 will receive standard care, wherein clinicians will follow national guidelines without AI assistance. Primary outcomes, assessed over 24 months, will include incidence of cardiovascular events, clinician patterns in decision-making (such as medication initiation or referrals), and patient-reported care satisfaction.

Main ethical concerns

Privacy, security and data management

Sharing de-identified health data across jurisdictions raises questions about regulatory alignment, control, and future use. Data management rules and regulations, such as in data ownership or in data use of AI technologies, for example, may vary depending on jurisdictions, which may be non-existent for some (77). Furthermore, third-party AI systems are always at risk of breaches, unauthorized access, and misuse of personal data (79). There is also the concern of advanced AI systems being able to reconstruct disparate anonymized data to reidentify patients. Since the large quantity of data needed to train AI systems requires cloud-based storage, the possibility of accessing data stored in the cloud by unauthorized agents leaves the sources of these data, even if anonymized, extremely vulnerable (80).

Autonomy and consent; transparency

Participants may not fully understand the meaning and implications of AI-based decision support, particularly with respect to uncertainty, limitations, and error. Traditional consent forms are difficult enough to understand; the additional burden of understanding AI-related concepts adds further complexity to obtaining informed consent. Transparency and clarity are essential to informed consent as they support understanding of risks and benefits. However, both are extremely challenging in the case of AI-driven trials due to the black box phenomenon wherein clinicians are unable to sufficiently explain how AI systems work (81). This makes it quite impossible for research participants to be fully informed. Furthermore, it is unclear where the responsibility of clinicians to inform patients ends and where the responsibility of patients to educate themselves begins as most users do not read user agreements, and monitoring as well as understanding the specifics software updates are extremely difficult (80).

Fairness and bias

The AI system may perform unevenly across sites due to differences in population characteristics, data completeness, and healthcare infrastructure. History of marginalization of several populations, for instance, can lead to bias in AI algorithms and, consequently, widen healthcare gaps and inequities (82). Furthermore, the scientific and institutional benefits of the trial may be more favorable towards those who are in high-income settings. Lack of exposure to AI-enabled technologies due to poor socioeconomic conditions contributes to a digital divide that renders those who do not have sufficient access to technology as well as those who have low digital literacy unable to maximize the benefits of AI in healthcare (82). Moreover, the accuracy of AI depends on the data sets that are used to train it, and thus, is not value-free (83). For this reason, AI algorithms that are trained on data from white men promote inequities and further reduces access to its social benefits. Finally, algorithmic risk scores, based on biased data, may influence clinical judgment that may lead to responsibility and liability concerns in the event of adverse outcomes. However, to whom responsibility falls for AI-driven errors resulting in harm, whether developers, clinicians, or health care institutions, is not clear cut (84). At the moment, there is no available framework that distinctly points at who should be responsible (and accountable) for AI use in healthcare (85).

Guidance

Privacy, security and data management

There must be rigorous legal safeguards that would ensure data protection, which must be respected and observed by developers of AI systems (80). Clear rules and agreements should define data ownership, access, and conditions for secondary use. Policymakers should take an active role in developing guidelines and infrastructures for creating more secure AI health systems (77). Oversight structures should include representatives from all participating jurisdictions with real decision-making authority, rather than merely consultative roles. More importantly, AI developers and healthcare institutions must work together to guarantee that patient data are stored securely in encrypted storage systems, access is restricted to authorized individuals, and data sources are secured (79).

Autonomy and consent; transparency

It is indispensable that sufficient information is given to users, patients, and even providers to be able to make informed decisions or change their minds in addition to establishing regulations for acquiring genuine informed consent (77). However, given the evolving role of AI during the trial, consent should be treated as an ongoing process rather than a one-time procedural requirement. This can be done by integrating AI into the clinical workflow (i.e., staff training, data management changes, and protocol updates), redesigning consent forms (i.e., adapted to local contexts and use accessible language), and transparent ethical disclosures (i.e., explain how the system works, what it cannot do, and where uncertainty remains) (81). As for the black box phenomenon, it is highly likely that a natural comprehension of AI systems will develop among healthcare practitioners as AI is increasingly integrated in their workflow, which will translate to a certain degree of understanding that will help them explain to patients what they need to know to arrive at informed decisions (80). This, evidently, can be ushered in sooner by conscious efforts among healthcare practitioners to study how these AI systems work.

Fairness and bias

Prior to the deployment of any AI system in health research, it should undergo site-specific validation. Across all AI development stages, model performance should be assessed across demographic groups to ensure consistent and optimal performance, and corrective steps should be taken before retraining and retesting algorithms when needed (82). This includes meticulous curation of data used to train AI systems (84). Performance metrics should be reported in stratified form, with clear thresholds that trigger review, recalibration, or suspension in cases of relevant and significant differences. Moreover, explicit and intentional inclusion of socioeconomic determinants in algorithmic models are vital in ensuring equity (86). Lastly, AI systems should be presented explicitly as decision support rather than as an authoritative decision-maker. Training on the use of AI systems in health research should emphasize that disagreement with algorithmic outputs is both expected and accepted and that clinical responsibility remains with practitioners, also known as the “human warranty” (80).

Case 3: AI-driven public health surveillance system

Introduction

Traditional public health surveillance relies on manual methods that heavily depend on systematic collection and analysis of data from various sources such as healthcare facilities, laboratories, and case reporting systems (87,88). As the volume of available data increases, traditional systems start struggling with “delayed reporting, under-detection cases, and limited scalability” (35,88,89). This has triggered the application of AI in epidemic intelligence.

AI and surveillance systems

Using machine learning, NLP, robotics, and blockchain (90), AI has “a transformative potential to revolutionize public health” (88). It can process and analyze huge amounts of data from various sources at scales far beyond human capacity. It can identify patterns in data that signal emerging public health threats. It helps to optimize resource allocation and facilitate real-time communication, especially during health crises (35,88,90). For example, on December 30, 2019, HealthMap raised a warning of “a cluster of cases of pneumonia of unknown cause” a few days before the first COVID-19 case was identified (91). In January 2020, BlueDot signaled pneumonia clusters in Wuhan days before the World Health Organization (WHO) public alert (21). ResNet-101 achieved the accuracy of 99.51% in distinguishing COVID-19 from other conditions. AI-driven technology was also used to analyze chest X-rays and computed tomography (CT) scans to detect COVID-19 with the promising accuracy of 84.3% to 100% (92). While AI makes public health surveillance systems more powerful, its limitations and ethical risks pose the question of how these systems can be deployed ethically and safely.

Case scenario

A Southeast Asian country has experienced several epidemics in the past three decades. To improve the country’s responses to future public health crises, its Department of Public Health (DPH) plans to commission the private vendor Guardisyn Inc. to develop an advanced AI-driven public health surveillance system. The new system aims to predict public health threats, trigger early investigation and rapid responses, and support effective decision-making. DPH will establish a new central office to host and manage the system and oversee its data collection, analysis, and outcomes. DPH would like the system to be ready for use in three years.

Data will be collected from the governmental health record database, hospital electronic records, environmental factors, and social media. Besides newly collected data, existing data will also be used for system training. The governmental health record database is used by the local public hospital network. It includes patients’ medical records, including medical diagnoses, allergies, and vaccination history. To get patients’ consent, the DPH plans to add an alert message to the database’s mobile application, which will pop up on patients’ smartphones if the application is installed, even if it is not being used. Through the DPH’s coordination, the local public and private hospitals will grant Guardisyn Inc. the access to their patient electronic records and laboratory databases. Patients will be asked to give consent when they arrive at medical appointments, such as doctor consultation, laboratory tests, and treatment procedures. In terms of environmental factors and social media, the DPH and Guardisyn Inc. have decided that the information is already publicly available and they do not need to bother the public for consent. Guardisyn Inc. and its employees involved in the project will sign the non-disclosure agreement to protect the privacy and safety of data sources.

Guardisyn Inc. plans to assign its best developer team on this project, aiming to deliver a strong and robust system to help the country with more rapid and effective responses to future public health emergencies. Guardisyn Inc. has also informed the DPH of the new system’s limitations, including the application of a “black box” algorithm. After the system is deployed, Guardisyn Inc. will continue providing technical support for another two years to ensure that the DPH is capable of operating the system and extracting important information on its own in five years from now.

Main ethical concerns

Privacy

Guardisyn Inc. plans to continue providing technical support for two years after the launch of the system. After this period, full responsibility for operation and maintenance will transition to the DPH. While this model is common for conventional information technology (IT) systems, concerns arise due to the AI system’s reliance on large-scale, continuously updated data streams. The system collects and processes sensitive health data, behavioral patterns, and potentially identifiable information from multiple sources. Over time, the AI model will evolve based on new data inputs, raising questions about long-term data governance and security. Recent industry findings (e.g., IBM) indicate that a significant proportion of organizations (more than one in ten) have experienced breaches involving AI systems, including attacks that exploit vulnerabilities in machine learning models themselves (93). Therefore, Guardisyn Inc. and the DPH should and must consider continuous post-implementation monitoring.

Autonomy & consent

The system requires access to patient medical records, which are obtained during hospital visits through standard consent forms. However, these forms are presented at the beginning of appointments, often when patients are stressed, unwell, or focused on receiving care. Is this the appropriate consent procedure? Do patients truly feel that they have the option to opt out without consequences to their care? What alternative consent models (e.g., dynamic consent, opt-in systems) could be more appropriate? Additionally, the system incorporates publicly available social media data to detect emerging health trends. While technically accessible, users may not expect their posts to be used for surveillance purposes. Is this an acceptable research protocol for this study? Finally, the system sends health alerts through a smartphone application. However, not all residents own smartphones, have installed the app, or have enabled notifications, raising concerns about unequal participation and implicit consent.

Transparency

The AI system relies on complex predictive models often described as “black box” algorithms. These models generate outbreak predictions and risk scores, but their internal decision-making processes are not easily interpretable by healthcare professionals or the public (94). Clinicians receiving alerts may not fully understand how the AI system reached its conclusions, yet they are expected to act on them. This creates tension between efficiency and accountability. Healthcare AI systems should be required to provide explainable outputs. But what level of transparency is necessary for public trust?

Fairness and bias

The system is trained on a combination of clinical data and social media inputs. However, these data sources may not fully represent the entire population. Certain groups, such as older adults, low-income populations, or individuals without internet access, may be underrepresented. Additionally, biases may emerge from the algorithm’s design or from patterns in historical data. If left unchecked, the system could disproportionately misclassify or overlook certain populations. How can developers ensure that training data are representative and inclusive? It is important to consider methods that should be used to detect and correct bias in AI systems. Further, another important point is the potential consequences if certain populations are systematically underdiagnosed or over-surveilled. Over time, as new data are incorporated, there is also a risk that existing biases may be reinforced or amplified (51).

Guidance

Privacy

While access to extensive datasets can enhance model performance and predictive accuracy, it also raises significant ethical concerns. In particular, the aggregation and processing of personal and sensitive information may expose individuals to privacy risks, especially given the probabilistic and generative nature of modern AI systems, which can inadvertently reveal or infer sensitive details (95). This creates a fundamental ethical tension: improving AI capabilities often depends on increased data access, yet greater data use may compromise individual privacy. Stakeholders must therefore carefully balance the potential clinical and public health benefits of AI with the obligation to protect personal data and uphold ethical standards (94,96,97).

Autonomy and consent

Another concern arises from the use of social media data without explicit user consent. Although such data may be publicly accessible, ethical use depends not only on accessibility but also on user expectations and intent. Individuals may share content for personal or social reasons without anticipating that it will be repurposed for surveillance or research. As highlighted by Williams et al. [2017], user expectations are a key factor in determining whether informed consent is required (98). Therefore, the use of social media data for AI-driven health monitoring should not be assumed to be ethically permissible solely based on its public availability.

Finally, issues of consent and inclusion emerge in relation to the system’s reliance on smartphone-based alerts. While the national health database may include records for all residents, not all individuals have equal access to smartphones, relevant applications, or notification settings. This creates disparities in who receives information and who is effectively included in the system’s benefits. As a result, reliance on digital delivery mechanisms may inadvertently exclude certain populations, raising concerns about equitable access and meaningful participation.

Transparency and accountability

When AI-driven systems produce outputs that cannot be interpreted or scrutinized, healthcare professionals may be unable to adequately evaluate their reliability or make informed clinical decisions. Researchers have emphasized the need for transparency and explainability in AI systems to ensure that they support rather than replace human judgment (3,28,40,44,99).

The European data protection regulation, General Data Protection Regulation (GDPR) emphasizes that personal data must be processed lawfully, fairly, and transparently in relation to the data subject (100). Similarly, the European Commission states that AI systems and their decisions should be explainable in a manner appropriate to the stakeholders involved (101).

These principles highlight that transparency is not merely a technical preference but an ethical and legal obligation. Ensuring explainability throughout the design, implementation, and deployment of AI systems is essential to maintain accountability, support clinical judgment, and build public trust. Regulatory frameworks at institutional, national, and international levels therefore play a critical role in enforcing transparency and safeguarding the responsible use of AI in healthcare.

More broadly, the ethical use of AI depends on shared accountability among all stakeholders. Responsibility does not rest solely with developers but extends to manufacturers, healthcare professionals, policymakers, and regulatory bodies. Each group plays a role in ensuring that AI systems operate in ways that uphold privacy, transparency, fairness, and ethical integrity. This includes the obligation to justify and explain how AI-generated decisions are made and applied in practice (51).

Fairness and bias

Bias is a well-recognized limitation of AI systems and can arise from both algorithm design and the data used for training. While efforts to ensure representative and high-quality data during the development phase are essential, they are not sufficient on their own. Bias can also emerge or be amplified during model deployment and real-world use. Therefore, stakeholders must remain vigilant not only during data collection but throughout the entire lifecycle of the system to prevent the reinforcement of existing inequalities. Unlike many traditional technical solutions, AI systems require ongoing evaluation, including regular audits to detect, measure, and mitigate bias as it evolves over time (51).

Given the current limitations of AI, human oversight remains indispensable. Effective oversight involves continuous monitoring to safeguard data privacy and security, routine audits to address bias, the establishment of appropriate regulatory frameworks at multiple levels, and clear communication and collaboration among stakeholders. As highlighted in the literature, AI should function as a tool to augment decision-making without undermining professional responsibility or individual autonomy (87,89,97).

Case 4: AI-based smartphone monitoring for early detection of depression relapse

Introduction

Depression is a common mental disorder involving a depressed mood or loss of pleasure or interest in activities for long periods of time. It affects approximately 332 million people worldwide (102). The impact of depression is significant; it is a leading cause of disability worldwide and contributes heavily to the global burden of disease (103). Depression is a highly recurrent illness, with relapse rates as high as 60% within two years of recovery. Early detection of relapse signals could enable timely interventions and reduce morbidity.

The problem of the high prevalence of depression and its recurrent nature is exacerbated by the gaps in mental health care across many countries globally. The lack of mental health care providers (104) and existing healthcare policies (105) are key barriers in the provision of care for those with mental diseases. In fact, it is estimated that more than 75% of people experiencing mental disorders in LMICs receive no treatment (103). The social stigma of mental disorders is common across cultures (106,107) and is a continued deterrent in seeking professional help. The use of emerging technology to address these issues could be effective in reducing depression in many contexts.

AI and mental health

Digital technologies, particularly AI, have significant potential to transform mental health care globally by improving access, early detection, and treatment delivery. AI-driven tools can analyze large datasets to identify patterns associated with mental health conditions, enabling earlier diagnosis and more personalized interventions (108). Collecting and pooling large datasets, including self-reported data, would generate more accurate predictive models in detecting symptoms that may require appropriate treatments as needed (109). Digital platforms such as mobile apps and telepsychiatry services could also help bridge gaps in care, especially in LMICs where mental health resources are scarce (110). Additionally, AI-powered chatbots and digital therapeutics can provide scalable, low-cost support for individuals with common conditions like depression and anxiety, reducing barriers related to stigma and workforce shortages (109,111). By expanding reach and enhancing efficiency, these technologies offer a promising pathway to more equitable and responsive mental health systems worldwide (112).

Case scenario

Investigators from the University of Thailand Medical Center (UMC) wanted to assess the feasibility and acceptability of continuous smartphone-based monitoring in adults with a history of major depressive disorder (MDD). Participants will be recruited through outpatient psychiatry clinics affiliated with the UMC. Integrated under Thailand’s “Gold Card” or Universal Coverage Scheme, services in the UMC-affiliated psychiatry clinics include psychiatric consultation and treatment, psychotropic medications and access to mental health hotlines. Clients aged 18–65 years with documented history of MDD but currently in remission will be asked to participate in the study during their regular visits to the clinic. Clients with current psychosis or other serious mental health conditions such as schizophrenia or bipolar disorder, or lack of smartphone access will be excluded.

Participants will complete the informed consent electronically via the MindTrack app, developed by MindWell, Inc., prior to enrollment. Details on the types of data to be collected, data-sharing arrangements with MindWell and UMC investigators, and the risks of data privacy breaches and limits of confidentiality will be provided. Provisions stating that the app is not a treatment tool nor a replacement to clinical care will be explicitly included in the consent. Participants may withdraw anytime from the study by uninstalling the app and the delivery of routine services for their care will not be withheld in the clinic.

Passive behavioral data streams [e.g., Global Positioning System (GPS), accelerometer, sleep/wake cycles, communication frequency] will be correlated with relapse episodes among the participants. Through the MindTrack app installed on their smartphones, the combined passive and self-reported data (e.g., daily mood ratings, weekly PHQ-9 surveys) will be used to validate the predictive AI/machine learning models for the study period of one year. If relapse indicators are detected by the app, participants will receive a notification suggesting they contact their health provider. The app will not provide direct clinical intervention, but participants will be reminded to continue routine care with their clinicians. The research findings could improve prevention of relapses through early detection of relapse indicators and reduce depression-related morbidity.

The MindTrack app will be developed by MindWell, Inc., a for-private company located in Bangkok. The app will be provided for free to participants for the duration of the study. Its parent company, Wellness Unlimited, Inc., is funding the study. MindWell will have access to de-identified datasets for future research and product development. Data will be stored primarily on MindWell’s secure servers, with secondary access for the study investigators from UMC. All research data collected will be coded by the team at MindWell before being shared with UMC investigators through a secure, password-protected research portal. Identifiers will not be collected (e.g., name, phone number, e-mail address), however, behavioral and location data are considered potentially re-identifiable. Encryption of data will be done in transit and at rest adhering to globally recognized and audited frameworks for information security management. Data will be unreadable to unauthorized users while moving across networks and while stored on devices to prevent data breaches during file transfers and physical theft. A data use agreement will be in place between UMC and MindWell. MindWell will provide technical support and infrastructure but has no role in clinical decision-making. UMC investigators will oversee independent data analysis to ensure scientific integrity. Any financial interests or advisory relationships of UMC investigators with MindWell, Inc. are disclosed.

Main ethical concerns

Privacy, security and data management

Though encryption of data will be done throughout the study from the collection, transfer and retention of the collated information, there will always be a risk for data breaches. Privacy must be vigilantly addressed as sensitive health information is collected. Beyond the risk of data breaches, the nature of passive data collection (e.g., GPS, communication logs, sleep patterns) introduces heightened concerns regarding re-identification and surveillance.

Additionally, the involvement of a private-sector partner (MindWell) raises concerns about secondary data use, commercialization, and data ownership. Participants may not fully anticipate how their data could be reused in future product development, creating ethical tension between innovation and respect for persons (113).

Autonomy and consent

Obtaining meaningful informed consent in AI-driven research is complex due to the technical opacity of machine learning systems. Participants may not fully understand how their data are processed, how predictions are generated, or the implications of algorithmic outputs (114). Moreover, the continuous and passive nature of monitoring challenges traditional models of consent, which are typically one-time events. Participants may not remain fully aware of ongoing data collection, raising concerns about ongoing autonomy and control. The potential psychological impact of receiving relapse alerts also raises questions about whether participants are adequately prepared for such notifications.

Fairness and bias

The full potential of apps in depression management is still evolving currently due to the limited scale of the sample sizes and the high risk of bias (109,112). AI learning models in identifying relapse indicators must address inequities by assuring a broad population group is represented in the study (115). AI systems trained on limited or non-representative datasets risk perpetuating systematic biases may potentially lead to inaccurate predictions for certain subgroups (52). In this study, recruitment from a specific clinical population (UMC-affiliated clinics) may limit generalizability and reinforce selection bias. Structural inequities, including differential smartphone usage patterns across socioeconomic groups, also affect data quality and model performance, raising concerns about fairness and bias (115,116).

Guidance

Privacy, security and data management

Even when traditional identifiers are removed, behavioral data can act as quasi-identifiers, making individuals potentially traceable (116). This is particularly sensitive in mental health research, where stigma associated with conditions like MDD may result in social, occupational, or even insurance-related harms if confidentiality is compromised. In addition to restricting access, researchers should adopt a data minimization principle, collecting only the data strictly necessary for study objectives.

Independent oversight bodies (e.g., data access committees) should regularly audit compliance. Cross-border data storage (e.g., cloud servers) may further complicate compliance with differing regulatory standards. Clear governance structures should define: (I) data ownership and stewardship responsibilities; (II) limits on secondary data use and commercialization; and (III) data retention timelines and secure deletion protocols. It is crucial to clearly define who has access to the data: access should be limited to authorized individuals with legitimate research purposes, governed by data use agreements, role-based permissions, and oversight from IRBs or data access committees. Researchers must also ensure that third-party collaborators, industry partners, and cloud service providers adhere to the same ethical and security standards. Transparency reports should be made available to participants, detailing how their data may be used over time.

Autonomy and consent

Full disclosure of the risks and benefits to participation in the study must be assured to the participants. The electronic collection of consent for participation must provide a robust process for the participants to clarify their concerns regarding the risks of participation. Consent processes should include a dynamic or ongoing consent model, allowing participants to revisit, modify, or withdraw consent as the study progresses. Digital platforms should include: (I) interactive consent modules (e.g., quizzes, layered explanations); (II) accessible summaries of AI functionality and limitations; and (III) clear explanations of potential risks, including potential psychological impacts of alerts. Participants should also be given granular control over data sharing (e.g., opting in/out of specific data streams like GPS). Participation must likewise be voluntary, with routine clinical services not withheld to them for non-participation in the study or withdrawal from the study.

Fairness and bias

Algorithmic bias in mental health contexts can have serious consequences, such as false positives (e.g., unnecessary anxiety) or false negatives (e.g., missed relapse detection). Accountability mechanisms—such as audit trails, documentation of data provenance, and ongoing monitoring—are essential to ensure responsible use. Continuous monitoring and recalibration of models should be implemented to prevent drift and emerging inequities.

Accountability should also extend to all stakeholders, including private partners. Transparency is critical: models should be accompanied by explainability mechanisms that allow researchers and clinicians to interpret predictions. Ethical AI governance frameworks, such as those proposed by the WHO [2024], should be incorporated to emphasize fairness, inclusiveness, transparency, and accountability as core principles (115). Further, it is also crucial to include diverse people with lived experience in the co-design of digital tools for mental health, as well as policy development to improve equitable solutions for all populations (117). Ultimately, ethical data stewardship in AI health research involves balancing innovation with respect for justice and the long-term societal implications of AI-driven health research.

Discussion

While the individual analysis of each case study has shown the complexities of using AI in health research, cross-examination of the aforementioned case studies reveals commonalities not only in ethical dilemmas but also in recommendations that can be concretely operationalized. Rather than treating ethical compliance as a checklist to be ticked or relying on reactive guidelines, researchers in AI, as well as research ethics committees that review AI-related health research protocols, need a general guide of recommendations for resolving ethical dilemmas that can be easily tailored to specific research protocols. Table 6 summarizes the core findings across the case studies and the actionable strategies that can be operationalized in AI-related health research.

Recommendations

The cases presented in this paper further demonstrate that ethical principles are fundamental to the responsible adoption of AI in healthcare, encompassing protection of patient privacy, safeguarding data security, mitigation of bias, necessity of transparency, and shared accountability. Based on the case studies, we recommend the checklist (Table 7), specifically for AI-assisted studies, for ethics review considerations.

The case studies have shown that countless research possibilities await in addressing ethical concerns for AI-enabled technologies. The ethical domains used can facilitate critical reflection for each but are inherently interrelated. Issues on privacy and safety may not be fully addressed if concerns on autonomy and consent are not complied with, for example. The myriad of combinations of these domains can lead to more innovative research questions on addressing ethical issues on AI-enabled health applications. Further, comprehensive guidance has been developed by the European Union, the WHO, AI4People, and other organizations addressing ethical principles for AI-driven health research, as well as broader regulatory and governance considerations. These resources provide important complementary perspectives and are recommended for further review (114,115).

Conclusions

In this paper we point out several key ethical issues of AI-enabled healthcare research: privacy and safety, autonomy and consent, transparency and accountability, and fairness and bias. Through four case studies, we provide a list of recommendations specifically for AI-enabled healthcare, beyond those for general healthcare research and practice. Based on these recommendations, we highlight that AI should augment instead of replacing human decision-making. Health researchers must appreciate that AI use in diagnosis, treatment and monitoring of mental health issues may broaden the landscape of digital mental health tools but should expect more ethical issues to emerge with the documented expanding capabilities of AI each year. The rising significance of big data, including how it will be collected, analyzed and used, must be detailed in future research. Our conclusion also emphasizes the close and continuous collaboration and communication between different stakeholders and their collective responsibility to justify and explain the decisions output by AI-systems.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Journal of Public Health and Emergency for the series “Towards Planetary Health Equity: A Global Call for Shared Solutions”. The article has undergone external peer review.

Peer Review File: Available at https://jphe.amegroups.com/article/view/10.21037/jphe-2026-0040/prf

Funding: This study received support from APRU, which sponsored the activity/paper from the grant received from Data.org to advance ‘AI and Data Science for Social Impact in Asia and the Pacific’. The funds were made available by the Mastercard Centre for Inclusive Growth.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jphe.amegroups.com/article/view/10.21037/jphe-2026-0040/coif). The series “Towards Planetary Health Equity: A Global Call for Shared Solutions” was commissioned by the editorial office without any funding or sponsorship. M.W. served as the unpaid Guest Editor of the series and serves as an Editor-in-Chief of Journal of Public Health and Emergency from July 2025 to June 2027. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This was not research with human subjects. Therefore, no ethical approvals were required.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Patel T, Ibrahim HY, Henna F, et al. A narrative review on artificial intelligence in neurosurgery: ethical challenges and implementation considerations. Ann Med Surg (Lond) 2025;87:8654-63. [Crossref] [PubMed]
2. Peiffer-Smadja N, Rawson TM, Ahmad R, et al. Machine learning for clinical decision support in infectious diseases: a narrative review of current applications. Clin Microbiol Infect 2020;26:584-95. Erratum in: Clin Microbiol Infect 2020 Aug;26(8):1118. [Crossref] [PubMed]
3. Mennella C, Maniscalco U, De Pietro G, et al. Ethical and regulatory challenges of AI technologies in healthcare: A narrative review. Heliyon 2024;10:e26297. [Crossref] [PubMed]
4. Harishbhai Tilala M, Kumar Chenchala P, Choppadandi A, et al. Ethical Considerations in the Use of Artificial Intelligence and Machine Learning in Health Care: A Comprehensive Review. Cureus 2024;16:e62443. [Crossref] [PubMed]
5. Steiner DF, MacDonald R, Liu Y, et al. Impact of Deep Learning Assistance on the Histopathologic Review of Lymph Nodes for Metastatic Breast Cancer. Am J Surg Pathol 2018;42:1636-46. [Crossref] [PubMed]
6. Wong D, Yip S. Machine learning classifies cancer. Nature 2018;555:446-7. [Crossref] [PubMed]
7. Mohan A, Asghar Z, Abid R, et al. Revolutionizing healthcare by use of artificial intelligence in esophageal carcinoma - a narrative review. Ann Med Surg (Lond) 2023;85:4920-7. [Crossref] [PubMed]
8. Lee S, Chu Y, Ryu J, et al. Artificial Intelligence for Detection of Cardiovascular-Related Diseases from Wearable Devices: A Systematic Review and Meta-Analysis. Yonsei Med J 2022;63:S93-S107. [Crossref] [PubMed]
9. Yasmin F, Shah SMI, Naeem A, et al. Artificial intelligence in the diagnosis and detection of heart failure: the past, present, and future. Rev Cardiovasc Med 2021;22:1095-113. [Crossref] [PubMed]
10. Elantary R, Othman S. Artificial Intelligence in Electrocardiography: From Automated Arrhythmia Detection to Predicting Hidden Cardiovascular Disease. Cureus 2025;17:e94065. [Crossref] [PubMed]
11. Lakhani P, Sundaram B. Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks. Radiology 2017;284:574-82. [Crossref] [PubMed]
12. Wong F, de la Fuente-Nunez C, Collins JJ. Leveraging artificial intelligence in the fight against infectious diseases. Science 2023;381:164-70. [Crossref] [PubMed]
13. Heidema S, Stoepker IV, Flaherty G, et al. From GeoSentinel data to epidemiological insights: a multidisciplinary effort towards artificial intelligence-supported detection of infectious disease outbreaks. J Travel Med 2024;31:taae013. [Crossref] [PubMed]
14. Sorrentino FS, Jurman G, De Nadai K, et al. Application of Artificial Intelligence in Targeting Retinal Diseases. Curr Drug Targets 2020;21:1208-15. [Crossref] [PubMed]
15. Firth J, Torous J, Nicholas J, et al. The efficacy of smartphone-based mental health interventions for depressive symptoms: a meta-analysis of randomized controlled trials. World Psychiatry 2017;16:287-98. [Crossref] [PubMed]
16. Bajaj S, Blair KS, Dobbertin M, et al. Machine learning based identification of structural brain alterations underlying suicide risk in adolescents. Discov Ment Health 2023;3:6. [Crossref] [PubMed]
17. Mansoor MA, Ansari KH. Early Detection of Mental Health Crises through Artifical-Intelligence-Powered Social Media Analysis: A Prospective Observational Study. J Pers Med 2024;14:958. [Crossref] [PubMed]
18. Bari LF, Ahmed I, Ahamed R, et al. Potential Use of Artificial Intelligence (AI) in Disaster Risk and Emergency Health Management: A Critical Appraisal on Environmental Health. Environ Health Insights 2023;17:11786302231217808. [Crossref] [PubMed]
19. Liu Z, Wang S, Zhang Y, et al. Artificial Intelligence in Food Safety: A Decade Review and Bibliometric Analysis. Foods 2023;12:1242. [Crossref] [PubMed]
20. Kutyauripo I, Rushambwa M, Chiwazi L. Artificial intelligence applications in the agrifood sectors. J Agric Food Res 2023;11:100502.
21. Chumachenko D, Kaur J, Chen JY. Editorial: Outbreak oracles: how AI's journey through COVID-19 shapes future epidemic strategy. Front Artif Intell 2025;8:1636444. [Crossref] [PubMed]
22. Batra M, Erbas B. Extreme Weather, Vulnerable Populations, and Mental Health: The Timely Role of AI Interventions. Int J Environ Res Public Health 2025;22:602. [Crossref] [PubMed]
23. Nonaka S, Fujii S, Yamada KC. Purposes of Smartphone Health Care Apps and the Practicality of Their Functions in Disaster Situations: Qualitative Function Assessment Study. JMIR Form Res 2025;9:e56862. [Crossref] [PubMed]
24. Tahernejad A, Sahebi A, Abadi ASS, et al. Application of artificial intelligence in triage in emergencies and disasters: a systematic review. BMC Public Health 2024;24:3203. [Crossref] [PubMed]
25. Rony MKK, Das A, Khalil MI, et al. The Role of Artificial Intelligence in Nursing Care: An Umbrella Review. Nurs Inq 2025;32:e70023. [Crossref] [PubMed]
26. Zaidan E, Ibrahim IA. AI governance in a complex and rapidly changing regulatory landscape: a global perspective. Humanit Soc Sci Commun 2024;11:1-18.
27. Palaniappan K, Lin EYT, Vogel S. Global Regulatory Frameworks for the Use of Artificial Intelligence (AI) in the Healthcare Services Sector. Healthcare (Basel) 2024;12:562. [Crossref] [PubMed]
28. El Arab RA, Al Moosa OA, Sagbakken M. Economic, ethical, and regulatory dimensions of artificial intelligence in healthcare: an integrative review. Front Public Health 2025;13:1617138. [Crossref] [PubMed]
29. Cao Y, Deng L, Liu X, et al. Ethical challenges in the algorithmic era: a systematic rapid review of risk insights and governance pathways for nursing predictive analytics and early warning systems. BMC Med Ethics 2025;26:151. [Crossref] [PubMed]
30. Elendu C, Amaechi DC, Elendu TC, et al. Ethical implications of AI and robotics in healthcare: A review. Medicine (Baltimore) 2023;102:e36671. [Crossref] [PubMed]
31. Grzybowski A, Jin K, Wu H. Challenges of artificial intelligence in medicine and dermatology. Clin Dermatol 2024;42:210-5. [Crossref] [PubMed]
32. Jobin A, Ienca M, Vayena E. The global landscape of AI ethics guidelines. Nat Mach Intell 2019;1:389-99.
33. Mohsin Khan M, Shah N, Shaikh N, et al. Towards secure and trusted AI in healthcare: A systematic review of emerging innovations and ethical challenges. Int J Med Inform 2025;195:105780. [Crossref] [PubMed]
34. Osnat B. Patient perspectives on artificial intelligence in healthcare: A global scoping review of benefits, ethical concerns, and implementation strategies. Int J Med Inform 2025;203:106007. [Crossref] [PubMed]
35. Villanueva-Miranda I, Xiao G, Xie Y. Artificial intelligence in early warning systems for infectious disease surveillance: a systematic review. Front Public Health 2025;13:1609615. [Crossref] [PubMed]
36. Hofmann B. Biases in AI: acknowledging and addressing the inevitable ethical issues. Front Digit Health 2025;7:1614105. [Crossref] [PubMed]
37. Achuthan K, Ramanathan S, Srinivas S, et al. Advancing cybersecurity and privacy with artificial intelligence: current trends and future research directions. Front Big Data 2024;7:1497535. [Crossref] [PubMed]
38. Murdoch B. Privacy and artificial intelligence: challenges for protecting health information in a new era. BMC Med Ethics 2021;22:122. [Crossref] [PubMed]
39. Prakash S, Balaji JN, Joshi A, et al. Ethical Conundrums in the Application of Artificial Intelligence (AI) in Healthcare-A Scoping Review of Reviews. J Pers Med 2022;12:1914. [Crossref] [PubMed]
40. Arjomandi Rad A, Vardanyan R, Athanasiou T, et al. The ethical considerations of integrating artificial intelligence into surgery: a review. Interdiscip Cardiovasc Thorac Surg 2025;40:ivae192. [Crossref] [PubMed]
41. Singhal A, Neveditsin N, Tanveer H, et al. Toward Fairness, Accountability, Transparency, and Ethics in AI for Social Media and Health Care: Scoping Review. JMIR Med Inform 2024;12:e50048. [Crossref] [PubMed]
42. Goktas P, Grzybowski A. Shaping the Future of Healthcare: Ethical Clinical Challenges and Pathways to Trustworthy AI. J Clin Med 2025;14:1605. [Crossref] [PubMed]
43. Gaffar S, Gearhart A. Artificial intelligence and wearable technology. Intelligence-Based Cardiology and Cardiac Surgery 2024:351-6.
44. Weiner EB, Dankwa-Mullan I, Nelson WA, et al. Ethical challenges and evolving strategies in the integration of artificial intelligence into clinical practice. PLOS Digit Health 2025;4:e0000810. [Crossref] [PubMed]
45. Laine J, Minkkinen M, Mantymaki M. Ethics-based AI auditing: a systematic literature review on conceptualizations of ethical principles and knowledge contributions to stakeholders. Inf Manage 2024;61: [Crossref]
46. Kenney A. (2025, June 19). How to instill AI ethics in your organization. Financial Management. Available online: https://www.fm-magazine.com/issues/2025/jun/how-to-instil-ai-ethics-in-your-organisation/47.
47. Adadi A, Berrada M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access 2018;6:52138-60.
48. Char DS, Shah NH, Magnus D. Implementing Machine Learning in Health Care - Addressing Ethical Challenges. N Engl J Med 2018;378:981-3. [Crossref] [PubMed]
49. Osasona F, Amoo OO, Atadoga A, et al. Reviewing the ethical implications of AI in decision making processes. Int J Manag Entrep Res 2024;6:322-35.
50. Ueda D, Kakinuma T, Fujita S, et al. Fairness of artificial intelligence in healthcare: review and recommendations. Jpn J Radiol 2024;42:3-15. [Crossref] [PubMed]
51. Radanliev P. Privacy, ethics, transparency, and accountability in AI systems for wearable devices. Front Digit Health 2025;7:1431246. [Crossref] [PubMed]
52. Obermeyer Z, Powers B, Vogeli C, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019;366:447-53. [Crossref] [PubMed]
53. Esmaeilzadeh P. Challenges and strategies for wide-scale artificial intelligence (AI) deployment in healthcare practices: A perspective for healthcare organizations. Artif Intell Med 2024;151:102861. [Crossref] [PubMed]
54. Abujaber AA, Nashwan AJ. Ethical framework for artificial intelligence in healthcare research: A path to integrity. World J Methodol 2024;14:94071. [Crossref] [PubMed]
55. Alzheimer's Disease International. Dementia statistics [Internet]. London: Alzheimer's Disease International; 2020 [cited 2026 Jan 25]. Available online: https://www.alzint.org/about/dementia-facts-figures/dementia-statistics/
56. Scheltens P, De Strooper B, Kivipelto M, et al. Alzheimer's disease. Lancet 2021;397:1577-90. [Crossref] [PubMed]
57. Babulal GM, Quiroz YT, Albensi BC, et al. Perspectives on ethnic and racial disparities in Alzheimer's disease and related dementias: Update and areas of immediate need. Alzheimers Dement 2019;15:292-312. [Crossref] [PubMed]
58. Kourtis LC, Regele OB, Wright JM, et al. Digital biomarkers for Alzheimer's disease: the mobile/ wearable devices opportunity. NPJ Digit Med 2019;2:9. [Crossref] [PubMed]
59. Gu Z, Ge B, Wang Y, et al. Artificial intelligence technologies for enhancing neurofunctionalities: a comprehensive review with applications in Alzheimer's disease research. Front Aging Neurosci 2025;17:1609063. [Crossref] [PubMed]
60. Ding Y, Sohn JH, Kawczynski MG, et al. A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using 18F-FDG PET of the Brain. Radiology 2019;290:456-64. [Crossref] [PubMed]
61. de la Fuente Garcia S, Ritchie CW, Luz S. Artificial Intelligence, Speech, and Language Processing Approaches to Monitoring Alzheimer's Disease: A Systematic Review. J Alzheimers Dis 2020;78:1547-74. [Crossref] [PubMed]
62. Javeed A, Dallora AL, Berglund JS, et al. Correction to: Machine Learning for Dementia Prediction: A Systematic Review and Future Research Directions. J Med Syst 2024;48:87. [Crossref] [PubMed]
63. Hui K, Ahn SS, Lye CT, et al. Ethical challenges of artificial intelligence in health care: a narrative review. Anesth Analg 2022;12:55-71.
64. Khurana D, Koli A, Khatter K, et al. Natural language processing: state of the art, current trends and challenges. Multimed Tools Appl 2023;82:3713-44. [Crossref] [PubMed]
65. Ohm P. Broken promises of privacy: responding to the surprising failure of anonymization. UCLA Law Rev 2012;57:1701-77.
66. Ford E, Pillinger S, Stewart R, et al. What is the patient re-identification risk from using de-identified clinical free text data for health research?. AI Ethics 2025;5:4441-54. [Crossref] [PubMed]
67. Fantus S, Li J, Wang T, et al. Ethical Knowledge, Challenges, and Institutional Strategies Among Medical AI Developers and Researchers: Focus Group Study. J Med Internet Res 2026;28:e79613. [Crossref] [PubMed]
68. Ficek J, Wang W, Chen H, et al. Differential privacy in health research: A scoping review. J Am Med Inform Assoc 2021;28:2269-76. [Crossref] [PubMed]
69. Sendak MP, Gao M, Brajer N, et al. Presenting machine learning model information to clinical end users with model facts labels. NPJ Digit Med 2020;3:41. [Crossref] [PubMed]
70. Farrar C. Trusted research environments (TREs)—a guide for beginners [Internet]. 2024 [cited 2026 Jan 27]. Available online: https://www.carnallfarrar.com/trusted-research-environments-guide-for-beginners/.
71. Ayankoya FY, Kuyoro SO, Nzenwata UJ, et al. A systematic review of machine learning and deep learning approaches for heart disease prediction. Asian J Res Comput Sci 2025;18:202-27.
72. Zhao D, Liu J, Xie W, et al. Cardiovascular risk assessment: a global perspective. Nat Rev Cardiol 2015;12:301-11. [Crossref] [PubMed]
73. Cai Y, Cai YQ, Tang LY, et al. Artificial intelligence in the risk prediction models of cardiovascular disease and development of an independent validation screening tool: a systematic review. BMC Med 2024;22:56. [Crossref] [PubMed]
74. Gul G, Korejo IA, Hakro DN, et al. Machine learning and ensemble methods for cardiovascular disease prediction: a systematic review of approaches, performance trends, and research challenges. Computers 2026;15:25.
75. Cai YQ, Gong DX, Tang LY, et al. Pitfalls in Developing Machine Learning Models for Predicting Cardiovascular Diseases: Challenge and Solutions. J Med Internet Res 2024;26:e47645. [Crossref] [PubMed]
76. Suri JS, Bhagawati M, Paul S, et al. Understanding the bias in machine learning systems for cardiovascular disease risk assessment: The first of its kind review. Comput Biol Med 2022;142:105204. [Crossref] [PubMed]
77. Mooghali M, Stroud AM, Yoo DW, et al. Trustworthy and ethical AI-enabled cardiovascular care: a rapid review. BMC Med Inform Decis Mak 2024;24:247. [Crossref] [PubMed]
78. Karabayir I, Akbilgic O. Generalizability of electrocardiographic artificial intelligence. NPJ Cardiovasc Health 2025;2:38. [Crossref] [PubMed]
79. Patel D, Chetarajupalli C, Khan S, et al. A narrative review on ethical considerations and challenges in AI-driven cardiology. Ann Med Surg (Lond) 2025;87:4152-64. [Crossref] [PubMed]
80. Lewin S, Chetty R, Ihdayhid AR, et al. Ethical Challenges and Opportunities in Applying Artificial Intelligence to Cardiovascular Medicine. Can J Cardiol 2024;40:1897-906. [Crossref] [PubMed]
81. Chau M, Rahman MG, Debnath T. From black box to clarity: Strategies for effective AI informed consent in healthcare. Artif Intell Med 2025;167:103169. [Crossref] [PubMed]
82. Mihan A, Pandey A, Van Spall HGC. Artificial intelligence bias in the prediction and detection of cardiovascular disease. NPJ Cardiovasc Health 2024;1:31. [Crossref] [PubMed]
83. Garcha I, Phillips SP. Social bias in artificial intelligence algorithms designed to improve cardiovascular risk assessment relative to the Framingham Risk Score: a protocol for a systematic review. BMJ Open 2023;13:e067638. [Crossref] [PubMed]
84. Chamsi-Pasha MAR, Chamsi-Pasha H. Ethical Challenges and Current Opportunities of Artificial Intelligence in Cardiology. Avicenna J Med 2025;15:95-9. [Crossref] [PubMed]
85. Bouhouita-Guermech S, Haidar H. Scoping Review Shows the Dynamics and Complexities Inherent to the Notion of "Responsibility" in Artificial Intelligence within the Healthcare Context. Asian Bioeth Rev 2024;16:315-44. [Crossref] [PubMed]
86. Varga TV. Algorithmic fairness in cardiovascular disease risk prediction: overcoming inequalities. Open Heart 2023;10:e002395. [Crossref] [PubMed]
87. Kaur J, Butt ZA. AI-driven epidemic intelligence: the future of outbreak detection and response. Front Artif Intell 2025;8:1645467. [Crossref] [PubMed]
88. Mendes VIS, Mendes BMF, Moura RP, et al. Harnessing artificial intelligence for enhanced public health surveillance: a narrative review. Front Public Health 2025;13:1601151. [Crossref] [PubMed]
89. Haykal D, Goldust M, Cartier H, et al. AI in humanitarian healthcare: a game changer for crisis response. Front Artif Intell 2025;8:1627773. [Crossref] [PubMed]
90. Lythreatis S, Acikgoz F, Yassine N. Artificial intelligence in humanitarian aid: A review and future research agenda. Technovation 2026;151:103415.
91. Parums DV. Editorial: Infectious Disease Surveillance Using Artificial Intelligence (AI) and its Role in Epidemic and Pandemic Preparedness. Med Sci Monit 2023;29:e941209. [Crossref] [PubMed]
92. Gawande MS, Zade N, Kumar P, et al. The role of artificial intelligence in pandemic responses: from epidemiological modeling to vaccine development. Mol Biomed 2025;6:1. [Crossref] [PubMed]
93. IBM. IBM report: 13% of organizations reported breaches of AI models or applications, 97% reported lacking proper AI access controls [Internet]. 2025 Jul 30 [cited 2026 Jan 25]. Available online: https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls.
94. Mahajan A, Heydari K, Powell D. Wearable AI to enhance patient safety and clinical decision-making. NPJ Digit Med 2025;8:176. [Crossref] [PubMed]
95. Stryker C, Scapicchio M. What is generative AI? [Internet]. IBM; [cited 2026 Jan 25]. Available online: https://www.ibm.com/think/topics/generative-ai.
96. Kenney A. How to instil AI ethics in your organization [Internet]. Financial Management. 2025 Jun 19 [cited 2026 Jan 25]. Available online: https://www.fm-magazine.com/issues/2025/jun/how-to-instil-ai-ethics-in-your-organisation/.
97. Piccialli F, di Cola VS, Giampaolo F, et al. The Role of Artificial Intelligence in Fighting the COVID-19 Pandemic. Inf Syst Front 2021;23:1467-97. [Crossref] [PubMed]
98. Williams ML, Burnap P, Sloan L, et al. Users' view of ethics in social media research: informed consent, anonymity, and harm. In: Woodfield K, editor. The ethics of online research. Bingley: Emerald Publishing Limited; 2017:27-52.
99. El Morr C, Ozdemir D, Asdaah Y, et al. AI-based epidemic and pandemic early warning systems: A systematic scoping review. Health Informatics J 2024;30:14604582241275844. [Crossref] [PubMed]
100. European Union. Article 5 of General Data Protection Regulation (GDPR) [Internet]. 2018 May 4 [cited 2026 Jan 25]. Available online: https://gdpr-info.eu/art-5-gdpr/.
101. European Commission. Ethics guidelines for trustworthy AI [Internet]. 2019 Apr 8 [cited 2026 Jan 25]. Available online: https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai.
102. World Health Organization (WHO). (2025). Depressive disorder (depression). Available online: https://www.who.int/news-room/fact-sheets/detail/depression.
103. World Health Organization (WHO). (2026). Depression. Available online: https://www.who.int/health-topics/depression.
104. McGINTY B. The Future of Public Mental Health: Challenges and Opportunities. Milbank Q 2023;101:532-51. [Crossref] [PubMed]
105. Rameez S, Nasir A. Barriers to mental health treatment in primary care practice in low- and middle-income countries in a post-covid era: A systematic review. J Family Med Prim Care 2023;12:1485-504. [Crossref] [PubMed]
106. Corrigan PW, Watson AC. Understanding the impact of stigma on people with mental illness. World Psychiatry 2002;1:16-20. [PubMed]
107. Ahad AA, Sanchez-Gonzalez M, Junquera P. Understanding and Addressing Mental Health Stigma Across Cultures for Improving Psychiatric Care: A Narrative Review. Cureus 2023;15:e39549. [Crossref] [PubMed]
108. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019;25:44-56. [Crossref] [PubMed]
109. Linardon J, Torous J, Firth J, et al. Current evidence on the efficacy of mental health smartphone apps for symptoms of depression and anxiety. A meta-analysis of 176 randomized controlled trials. World Psychiatry 2024;23:139-49. [Crossref] [PubMed]
110. Murphy JK, Khan A, Sun Q, et al. Needs, gaps and opportunities for standard and e-mental health care among at-risk populations in the Asia Pacific in the context of COVID-19: a rapid scoping review. Int J Equity Health 2021;20:161. [Crossref] [PubMed]
111. Fitzpatrick KK, Darcy A, Vierhile M. Delivering Cognitive Behavior Therapy to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot): A Randomized Controlled Trial. JMIR Ment Health 2017;4:e19. [Crossref] [PubMed]
112. Myers A, Chesebrough L, Hu R, et al. Evaluating Commercially Available Mobile Apps for Depression Self-Management. AMIA Annu Symp Proc 2020;2020:906-14. [PubMed]
113. Mittelstadt BD, Allo P, Taddeo M, et al. The ethics of algorithms: Mapping the debate. Big Data Soc 2016;3:2053951716679679.
114. Floridi L, Cowls J, Beltrametti M, et al. AI4People-An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations. Minds Mach (Dordr) 2018;28:689-707. [Crossref] [PubMed]
115. World Health Organization (WHO). (2021). Ethics and governance of artificial intelligence for health. Available online: https://iris.who.int/server/api/core/bitstreams/f780d926-4ae3-42ce-a6d6-e898a5562621/content.
116. Seastedt KP, Schwab P, O'Brien Z, et al. Global healthcare fairness: We should be sharing more, not less, data. PLOS Digit Health 2022;1:e0000102. [Crossref] [PubMed]
117. Murphy JK, Saker S, Ananyo Chakraborty P, et al. Advancing equitable access to digital mental health in the Asia-Pacific region in the context of the COVID-19 pandemic and beyond: A modified Delphi consensus study. PLOS Glob Public Health 2024;4:e0002661. [Crossref] [PubMed]