The use of machine learning, or “artificial intelligence” (AI) in medicine is widespread and growing. In this paper, I focus on a specific proposed clinical application of AI: using models to predict incapacitated patients’ treatment preferences. Drawing on results from machine learning, I argue this proposal faces a special moral problem. Machine learning researchers owe us assurance on this front before experimental research can proceed. In my conclusion I connect this concern to broader issues in AI safety.

BackgroundHealthcare providers have to make ethically complex clinical decisions which may be a source of stress. Researchers have recently introduced Artificial Intelligence (AI)-based applications to assist in clinical ethical decision-making. However, the use of such tools is controversial. This review aims to provide a comprehensive overview of the reasons given in the academic literature for and against their use.MethodsPubMed, Web of Science, Philpapers.org and Google Scholar were searched for all relevant publications. The resulting set of publications was title and abstract (...) screened according to defined inclusion and exclusion criteria, resulting in 44 papers whose full texts were analysed using the Kuckartz method of qualitative text analysis.ResultsArtificial Intelligence might increase patient autonomy by improving the accuracy of predictions and allowing patients to receive their preferred treatment. It is thought to increase beneficence by providing reliable information, thereby, supporting surrogate decision-making. Some authors fear that reducing ethical decision-making to statistical correlations may limit autonomy. Others argue that AI may not be able to replicate the process of ethical deliberation because it lacks human characteristics. Concerns have been raised about issues of justice, as AI may replicate existing biases in the decision-making process.ConclusionsThe prospective benefits of using AI in clinical ethical decision-making are manifold, but its development and use should be undertaken carefully to avoid ethical pitfalls. Several issues that are central to the discussion of Clinical Decision Support Systems, such as justice, explicability or human–machine interaction, have been neglected in the debate on AI for clinical ethics so far.Trial registrationThis review is registered at Open Science Framework (https://osf.io/wvcs9). (shrink)

The modern legal and ethical movement against traditional welfare paternalism in medical decision-making extends to how decisions are made for patients lacking decisional capacity, prioritising surrogates’ judgment about what patients would have decided over even their best interests. In England and Wales, the Mental Capacity Act 2005 follows this trend of prioritising the patient’s prior wishes, values and beliefs but the dominant interpretation in life-sustaining treatment cases does so by in effect calling those values the ‘best interests’ of the patient (...) and focusing nearly exclusively on the ‘subjective’ viewpoint of the patient. In this article, we examine the recent Court of Protection judgment inBarnsley Hospitals NHS Foundation Trust v MSP[2020] EWCOP 26, which adhered closely to this approach, to suggest that it could have unexpected negative consequences. These include insufficient information gathering about and attention to patients’ objective medical interests, inadequacy of the evidentiary standard used for the substituted decision-making and, in some cases, even prioritising a surrogate’s current substituted judgment over the potential for an actual judgment by the patient. (shrink)

The patient preference predictor is a proposed computer-based algorithm that would predict the treatment preferences of decisionally incapacitated patients. Incorporation of a PPP into the decision-making process has the potential to improve implementation of the substituted judgement standard by providing more accurate predictions of patients’ treatment preferences than reliance on surrogates alone. Yet, critics argue that methods for making treatment decisions for incapacitated patients should be judged on a number of factors beyond simply providing them with the treatments they would (...) have chosen for themselves. These factors include the extent to which the decision-making process recognises patients’ freedom to choose and relies on evidence the patient themselves would take into account when making treatment decisions. These critics conclude that use of a PPP should be rejected on the grounds that it is inconsistent with these factors, especially as they relate to proper respect for patient autonomy. In this paper, we review and evaluate these criticisms. We argue that they do not provide reason to reject use of a PPP, thus supporting efforts to develop a full-scale PPP and to evaluate it in practice. There are no data in this work. (shrink)

When making substituted judgments for incapacitated patients, surrogates often struggle to guess what the patient would want if they had capacity. Surrogates may also agonize over having the (sole) responsibility of making such a determination. To address such concerns, a Patient Preference Predictor (PPP) has been proposed that would use an algorithm to infer the treatment preferences of individual patients from population-level data about the known preferences of people with similar demographic characteristics. However, critics have suggested that even if such (...) a PPP were more accurate, on average, than human surrogates in identifying patient preferences, the proposed algorithm would nevertheless fail to respect the patient’s (former) autonomy since it draws on the ‘wrong’ kind of data: namely, data that are not specific to the individual patient and which therefore may not reflect their actual values, or their reasons for having the preferences they do. Taking such criticisms on board, we here propose a new approach: the Personalized Patient Preference Predictor (P4). The P4 is based on recent advances in machine learning, which allow technologies including large language models to be more cheaply and efficiently ‘fine-tuned’ on person-specific data. The P4, unlike the PPP, would be able to infer an individual patient’s preferences from material (e.g., prior treatment decisions) that is in fact specific to them. Thus, we argue, in addition to being potentially more accurate at the individual level than the previously proposed PPP, the predictions of a P4 would also more directly reflect each patient’s own reasons and values. In this article, we review recent discoveries in artificial intelligence research that suggest a P4 is technically feasible, and argue that, if it is developed and appropriately deployed, it should assuage some of the main autonomy-based concerns of critics of the original PPP. We then consider various objections to our proposal and offer some tentative replies. (shrink)