Forty-four publications were analyzed in this systematic review. A complete list can be found in “Publications included in additional files”. The sample consists mainly of magazine articles (n= 24) or commentary (n= 20), primarily the American Journal of Bioethics (n= 16), Journal of Medical Ethics (n= 10) and the Journal of Medicine and Philosophy (n= 9). One of the papers is based on empirical research. The first authors of the analyzed publications were primarily affiliated with US institutions (n= 28) and Europe (n= 14), and there are only two in South Africa (n= 1) and Korea (n= 1). Figure 2 shows the year profile of papers, confirming the relatively recent emergence of the topic. 2014 and his cluster around 2022 are related to the first article on Patient Preference Predictor and his METHAD, and most of the included publications mention this article. Characteristics of these and other AI tools described in the sample are shown in Table 3.
Number of publications included in the year
The extracted reasons are presented below in relation to the ethical principles to which they relate. We focus on the most prominent reasons that have played the greatest role in the ethical debate about AI in clinical ethical decision-making. All identified reasons and their frequency of occurrence are specified in the “Additional File Code List” file.
A key outcome of the analysis is the fact that the scientific community has high hopes for AI for enhancing autonomy in clinical decision-making. Proposed AI application “will yield more accurate predictions than existing methods” [26] It “increases the likelihood that incapacitated patients will receive the care they want and avoid the care they don’t want.” [27]. This is considered particularly important in many cases where there is a lack of adequate advance directives available. This is because the alternative strategy of proxy-assisted decision-making “often fails to provide treatment that meets the patient’s wishes.” [28].Artificial intelligence tools are also seen as having ‘the potential to improve the transparency of ethical decision-making’ [29] thereby improving proxy decision-making [28, 30]enabling new lines of action for clinicians [31]and strengthen respect for the autonomy of all stakeholders in the process. [32, 33]. Conversely, some ethicists fear the opposite. That is, reducing the process of ethical deliberation to the statistical correlations found in the training data underlying ML tools. [27, 34]thus, deploying social and demographic characteristics as the sole determinants of their preferences, AI “could potentially jeopardize patient autonomy.” [35]. Artificial intelligence can contradict stakeholder intuition by providing supporting information based on faulty assumptions. [28, 36, 37]or less accurate than the surrogate prediction [29, 38] (Critics believe this is likely because “preferred unpredictable volatility inherently limits any forecasting model.”) [38]).
It is also hypothesized that AI may improve benefits obtained through clinical ethical decision-making by providing reliable information, such as “providing evidence of patient preference.” [39], to support clinicians, surrogate mothers, or other interested parties. This will benefit the overall quality of clinical care and the overall decision-making process. [24, 28].Additional reassurance may ‘help alleviate some of the burden associated with decision-making for helpless patients’ [23] Especially for surrogate mothers (as well as health care professionals and clinical ethicists). This burdensome situation occurs not only in large clinics where CESS is readily available, but also in small hospitals and primary care. Equipping ethics AI as a tool that is easily accessible to any digital terminal device will enable this kind of ethics support for a wider range of users and situations. [26, 40]potentially saving human and financial resources in the health system [26, 41]. Moreover, the use of AI in clinical settings may provide a form of cognitive moral enhancement. [42] promote ethical competence [43]. On the other hand, critics argue that the information provided by AI may not be as robust as thought, as “even well-performing algorithms may be unreliable in individual cases.” claims that there is [24]. The algorithms on which the tools are based may not fully cover the actual ethical decision-making process, as complex deliberations “are unlikely to be successfully reduced to a set of equations.” [43].Additionally, AI is believed to be incapable of acting empathically. [40] Or consider structural and systematic knowledge, because “algorithms still struggle to grasp context and explanations” [24].If that’s true, there may be no clear benefit to using AI [41]but may still impair the ability of stakeholders [44].
Artificial intelligence could help in terms of non-maliciousness by adding to traditional methods of ethical decision-making, but would otherwise “significantly stress surrogate decision-makers.” .[28]. This support could come in the form of advice to inform and assist stakeholders who “may need to make high-stakes decisions.”[39]and “often difficult to distinguish between patient preferences and own preferences.” [45]. On the contrary, some authors are concerned that the application in question “may increase the stress of some surrogate decision makers.” [32] by undermining trust when the AI disagrees with its assessment of the situation [39]or by reinforcing cognitive biases and deficiencies in decision making [46]. Another “disappointing consequence is that the development of ethical sensitivities that can be gained from working with difficult ethical cases is likely to be limited.”[31]which can lead to skill deterioration.
Some authors hope AI will reduce bias and increase justice and fairness in clinical decision-making. [25] and “offer”[ing] Objective, accurate and individualized assessment of difficult cases.” [25]. That said, most of the extracted reasons point in a different direction. As a direct result of their respective training data choices, AI usage may “simply reflect pre-existing biases.” [26] And thus “social injustice perpetuates”. [47]. Additionally, healthcare organizations (hospitals, health insurers, industry) may introduce new biases “to make projections favorable to hospital budgets.” [32].
Explainability has received less weight in ethical debates than other principles.Reasons mentioned focused on lack of transparency [35]explainability [35] Accountability in AI-supported decision-making [31]. “Concerns have been expressed about trusting ‘black boxes’,” while [24]most of the reviewed publications did not mention relevant elements.
Several reasons did arise for non-compliance with biomedical principles, most of which focused directly on deficiencies in the development of advisory applications.Problems with AI developed based on faulty assumptions and lack of data [37, 47]too many resources to develop [48] Or, “Automating ethical decision-making with AI is currently not feasible or ethical.” [49] It is cited as a reason for rejecting implementation. Moreover, proxy decision-making itself is described as superior to AI alternatives. [38].
